ROx3: RETINAL OXIMETRY UTILIZING THE BLUE

ROx3: RETINAL OXIMETRY UTILIZING THE BLUE
ROx3: RETINAL OXIMETRY UTILIZING THE BLUE-GREEN OXIMETRY
METHOD
by
Jennifer Kathleen Hendryx Parsons
_____________________________
Copyright © Jennifer Kathleen Hendryx Parsons 2014
A Dissertation Submitted to the Faculty of the
DEPARTMENT OF OPTICAL SCIENCES
In Partial Fulfillment of the Requirements
For the Degree of
DOCTOR OF PHILOSOPHY
In the Graduate College
THE UNIVERSITY OF ARIZONA
2014
THE UNIVERSITY OF ARIZONA
GRADUATE COLLEGE
As members of the Dissertation Committee, we certify that we have read the dissertation
prepared by Jennifer K. Hendryx Parsons, titled ROx3: Retinal Oximetry Utilizing the
Blue-Green Oximetry Method and recommend that it be accepted as fulfilling the
dissertation requirement for the Degree of Doctor of Philosophy.
_______________________________________________ Date: 11/10/2014
Dr. Russell A. Chipman
_______________________________________________ Date: 11/10/2014
Dr. Kurt R. Denninghoff, MD
_______________________________________________ Date: 11/10/2014
Dr. Arthur F. Gmitro
Final approval and acceptance of this dissertation is contingent upon the candidate’s
submission of the final copies of the dissertation to the Graduate College.
I hereby certify that I have read this dissertation prepared under my direction and
recommend that it be accepted as fulfilling the dissertation requirement.
________________________________________________ Date: 11/10/2014
Dissertation Director: Dr. Russell A. Chipman
2
STATEMENT BY AUTHOR
This dissertation has been submitted in partial fulfillment of the requirements for
an advanced degree at the University of Arizona and is deposited in the University
Library to be made available to borrowers under rules of the Library.
Brief quotations from this dissertation are allowable without special permission,
provided that an accurate acknowledgement of the source is made. Requests for
permission for extended quotation from or reproduction of this manuscript in whole or in
part may be granted by the copyright holder.
SIGNED: Jennifer K. Hendryx Parsons
3
ACKNOWLEDGEMENTS
Ascribe to the LORD the glory due His name; worship the LORD in the splendor
of His holiness. Psalm 29:2
I could fill these pages with acknowledgements alone; I am incredibly grateful for
the support I have received from many dear friends and family. I appreciate all those
whose generosity and hospitality made it possible to finish this degree as a distance
student. I especially want to thank my parents, Forrest and Rebecca Hendryx, for their
example of work ethic; in addition to my sister, Emily Hendryx, they have been a
perpetual source of love and encouragement throughout my entire life.
This dissertation would not be possible without the knowledge, guidance, and
mentorship of Dr. Russell Chipman and Dr. Kurt Denninghoff. They are each such a
wealth of knowledge and experience, and I have appreciated the instruction and candid
advice I have received from these gentlemen. It has truly been an honor to learn from
and work with them.
The progress achieved on the ROx has been the product of teamwork between
Tyson Ririe, Kasia Sieluzycka, and myself. Tyson has been a friend and invaluable
partner in the entirety of this work. His perspectives, questions, ideas, and efforts have
been instrumental in the development of the results and discussions in this dissertation.
Another driving force in this work has been Dr. Lawrence DeLuca. Larry’s energy,
enthusiasm, working knowledge, and skill have propelled the pig experiments that are
such a significant part of this dissertation, and his friendship has been a frequent source
of motivation and encouragement. I also appreciate the work of undergraduate and
medical students who spent time on this project: Benjamin Juan, Fernando Tapia,
Kristina Voss, Brendan Munzer, and David Liu.
I also owe a debt of gratitude to Dr. Arthur Gmitro. Art spearheaded the BMIS
fellowship program that ultimately funded me through the majority of my graduate
career, as well as exposed me to other areas of biomedical research at the University of
Arizona.
Finally, I need to express my deep gratitude to my loving husband, Douglas
Gerald Parsons. His patience, support, and sense of humor have been instrumental in the
completion of this work and my sanity in the mean time.
4
Table of Contents
TABLE OF FIGURES ...................................................................................................... 9
TABLE OF TABLES...................................................................................................... 21
ABSTRACT ..................................................................................................................... 22
Chapter 1: INTRODUCTION ...................................................................................... 23
A.
Introduction ........................................................................................................ 23
B.
What is Retinal Oximetry? ................................................................................. 23
C.
Description of Content ....................................................................................... 24
D.
Summary ............................................................................................................ 25
Chapter 2: PHYSIOLOGICAL BACKGROUND ...................................................... 27
A.
Introduction ........................................................................................................ 27
B.
Blood .................................................................................................................. 27
C.
Oxygen Transport ............................................................................................... 30
D.
The Eye .............................................................................................................. 32
a.
Anterior Chamber ............................................................................................... 32
b.
The Crystalline Lens .......................................................................................... 33
c.
Posterior Chamber .............................................................................................. 34
d.
The Retina .......................................................................................................... 34
e.
Vasculature of the Eye ....................................................................................... 36
f.
Motion of the Eye ............................................................................................... 37
E. Optical Layout of the Human Eye ......................................................................... 38
F.
Conclusions ............................................................................................................ 41
Chapter 3: HISTORY AND PRINCIPLES OF RETINAL OXIMETRY ................ 43
A.
Introduction ........................................................................................................ 43
B.
History of Oximetry ........................................................................................... 43
C.
Principles of Oximetry ....................................................................................... 44
a.
Lambert-Beer Law ............................................................................................. 45
b.
Scattering ............................................................................................................ 47
D.
Retinal Oximetry ................................................................................................ 49
E. Blue-Green Oximetry............................................................................................. 52
5
Chapter 4: DEVICE DESCRIPTION........................................................................... 57
A.
Introduction ........................................................................................................ 57
B.
Apparatus Requirements .................................................................................... 57
C.
Optics Overview ................................................................................................. 64
a.
ROx Light Path Summary .................................................................................. 64
b.
Details................................................................................................................. 68
D.
Scanning System ................................................................................................ 76
E. Power Delivery ...................................................................................................... 86
F.
Vessel Targeting and Identification ....................................................................... 90
G.
Summary ............................................................................................................ 94
Chapter 5: CALIBRATION EXPERIMENTS ON LIVE SWINE ............................ 96
A.
Introduction ........................................................................................................ 96
B.
Trial Run ............................................................................................................ 96
C.
First Calibration Experiment ............................................................................ 102
a.
Setup ................................................................................................................. 102
b.
Experimental Procedure ................................................................................... 107
D.
Calibration Data Analysis ................................................................................ 108
E. Vessel Diameter Measurements ........................................................................... 119
F.
Conclusions from First Calibration Experiment .................................................. 121
G.
Second Calibration Experiment ....................................................................... 122
a.
Setup and Procedure ......................................................................................... 123
b.
Data and Analysis............................................................................................. 127
c.
Conclusions from Second Calibration Experiment .......................................... 129
H.
Conclusions ...................................................................................................... 130
Chapter 6: SEPSIS EXPERIMENT ON LIVE SWINE ........................................... 132
A.
Introduction ...................................................................................................... 132
B.
Sepsis and the ROx .......................................................................................... 132
C.
Sepsis Experiment ............................................................................................ 134
a.
Setup ................................................................................................................. 135
b.
Surgical Procedure ........................................................................................... 138
c.
ROx Data Acquisition ...................................................................................... 142
6
D.
Results .............................................................................................................. 144
a.
Sepsis Model .................................................................................................... 144
b.
ROx Data .......................................................................................................... 149
E. Discussion ............................................................................................................ 158
a.
Sepsis Model .................................................................................................... 158
b.
ROx Data .......................................................................................................... 160
F.
Conclusions .......................................................................................................... 167
Chapter 7: HUMAN EYE EXPERIMENTS .............................................................. 169
A.
Introduction ...................................................................................................... 169
B.
Considerations for Human Data Acquisition ................................................... 169
a.
Patient Comfort ................................................................................................ 169
b.
Practicality for Clinical User ............................................................................ 173
C.
Experimental Setup .......................................................................................... 176
D.
Results .............................................................................................................. 180
a.
March 26, 2014 ................................................................................................ 180
b.
May 8, 2014...................................................................................................... 185
c.
June 20, 2014.................................................................................................... 189
d.
June 23, 2014.................................................................................................... 192
e.
June 25, 2014.................................................................................................... 197
E. Conclusions .......................................................................................................... 199
Chapter 8: DATA ACQUISITION AND ANALYSIS METHODS ......................... 201
A.
Introduction ...................................................................................................... 201
B.
Image Preparation ............................................................................................ 201
a.
Frame Grabber Image Acquisition and Preparation ......................................... 201
b.
PicoScope Data Acquisition ............................................................................. 207
c.
PicoScope Image Preparation from Sub-sampled Data ................................... 210
d.
PicoScope Image Preparation from Complete Data Set................................... 214
C.
Vessel and Spectrally Neutral Fundus Identification Methods ........................ 217
D.
Finding the Optical Density of the Vessel ....................................................... 228
a.
Theoretical Vessel Profile ................................................................................ 229
b.
Vessel Fitting by Hand ..................................................................................... 231
7
c.
Automated Vessel Fitting ................................................................................. 232
d.
Iterative Process of Automated Fitting............................................................. 238
E. Determining SO2 From the Vessel Images .......................................................... 244
a.
Error Analysis .................................................................................................. 244
b.
Error Analysis for Future Consideration .......................................................... 245
F.
Conclusions .......................................................................................................... 247
Chapter 9: MINIATURIZATION OF THE ROX..................................................... 250
A.
Introduction ...................................................................................................... 250
B.
Requirements for the ROx as a Medical Device .............................................. 250
C.
Miniaturized Optical Design ............................................................................ 251
a.
Design Process: Illumination Path ................................................................... 259
b.
Design Process: Focusing Lenses .................................................................... 267
c.
Design Process: Return Path ............................................................................ 274
d.
Design Process: Mounting Considerations ...................................................... 277
D.
Further System Miniaturization ....................................................................... 277
E. Conclusions .......................................................................................................... 279
Chapter 10: DISCUSSION AND CONCLUSIONS................................................... 280
A.
Introduction ...................................................................................................... 280
B.
Image Quality ................................................................................................... 280
a.
Further Glint Analysis and Prevention ............................................................. 281
b.
Ghost Reflections from Lenses ........................................................................ 289
C.
System and Analysis Automation .................................................................... 291
a.
System Automation .......................................................................................... 291
b.
Automated Image Analysis Procedure ............................................................. 293
c.
Other Considerations ........................................................................................ 296
D.
In Vivo Experiments ......................................................................................... 299
E. Summary .............................................................................................................. 301
REFERENCES.............................................................................................................. 302
8
TABLE OF FIGURES
Figure 2.1: Image of a typical human left eye. The encircled region on the left
corresponds to the ONH, the outer encircled region on the right corresponds to the
fovea, and inner encircled region on the right corresponds to the foveola. This
photograph was taken by the fundus camera in during a routine eye exam
(Permission, Douglas Parsons). .............................................................................36
Figure 2.2: Optical layout of the eye (front nodal point labeled as N). The Arizona
Eye Model (relaxed) is shown as rendered by Code V. .........................................40
Figure 3.1: Optical density spectra of hemoglobin (Hb) and oxyhemoglobin
(HbO2)45. ................................................................................................................52
Figure 3.2: Three light paths of interest when a retinal vessel is illuminated by a
converging cone of light: (a) the glint, which reflects directly from the surface of
the vessel and conveys little information about the oxygen saturation, and (b)
scattered light that has interacted with red blood cells, either in transmission or in
backscatter..............................................................................................................53
Figure 4.1: Diagram of the unfolded basic optical system of the ROx. Starting from
the pellicle, the forward beam path is shown as a solid line. The scan mirrors path
shows how the FSM, SSM, and iris are conjugate. The scanning beam paths
show the behavior of the beam at maximum scan angles. The return path and the
forward path are counter propagating until the pellicle; the long dashes show the
return path after the pellicle to the PMT. The diagram is not to scale. “OAP”
denotes off-axis paraboloids. .................................................................................66
Figure 4.2: Code V diagram of the return light path (compare to Figure 4.4). The
centered blue ray bundle is the reflected light path when the scan mirrors are
centered. The green bundle comes from a retinal point 250μm from the axis when
the scan mirrors are centered. ................................................................................70
Figure 4.3: Code V diagram of the forward light path (compare to Figure 4.4). The
center blue ray bundle is the light path when the scan mirrors are centered. The
outer red and green bundles are the light paths when the scan mirrors are at - and
+ 3 degrees, respectively, corresponding to a scan size of ~500 x 500 μm. ..........71
Figure 4.4: Photo of optical layout. The blue arrow into the AOTF indicates the Ar++
multispectral beam, and the green arrow out of the AOTF indicates the beam of a
single wavelength selected by the AOTF. The red arrows indicate the path of the
IR targeting beam, which then becomes collinear with the light path from the
AOTF at the cold mirror. The gold arrow indicates the return path detected by the
PMT. Note that any cardboard serves to contain stray light behind the pinhole
filter which is hidden in the photograph. ...............................................................72
Figure 4.5: Comparison or retinal scans when there is no glint (left) and when the
glint is present as in increased intensity (right). These images were gathered from
9
an enucleated swine eye, and it was not possible to determine whether this is a
vein or artery. .........................................................................................................74
Figure 4.6: Raw interlaced image with edge blanking. Image doubling is occurring
in the fast scan orientation (the horizontal is the fast scan and the vertical is the
slow scan). The magnified image makes the individual wavelengths and blank
lines apparent. ........................................................................................................77
Figure 4.7: Signals corresponding to the SSM; the SSM angle (examples shown at
top) is controlled by the voltage from the signal generator (in black). ..................77
Figure 4.8: The image in Figure 4.6 is separated into the 6 channels. Because each
wavelength is used after every sixth line, the raw images are not equally
proportioned in the horizontal and vertical directions. It is evident that the edges
are not blanked perfectly. The top 5 are false-color images to make the contrast
more clear...............................................................................................................78
Figure 4.9: The 514.5nm image from Figure 4.6 with equal horizontal and vertical
proportions. ............................................................................................................78
Figure 4.10: Signals corresponding to the FSM .........................................................80
Figure 4.11: Photo of improved optical layout. The blue arrow into the AOTF
indicates the Ar++ multispectral beam, and the green arrow out of the AOTF
indicates the beam of a single wavelength selected by the AOTF, with the dashed
green line indicating the reference path to the reference PMT. The red arrows
indicate the IR targeting path, which becomes collinear with the light bath from
the AOTF at the cold mirror. The gold arrow indicates the return path to the
primary PMT. Note that the cardboard contains stray light behind the pinhole
filter. .......................................................................................................................83
Figure 4.12: Specs provided by JDSU on the Ar++ laser used by the ROx. The graph
on the left shows how the power of each wavelength increases with current. The
graph on the right shows the percentage of which each wavelength comprises the
total beam for a given power. The current used in eye data acquisition is 7.96 A.88
Figure 4.13: Data taken by my colleague Tyson Ririe, showing the intensity from the
pellicle in reflection as the angle and wavelength are varied. ...............................89
Figure 4.14: Transmission and reflection spectra of the pellicle at 27°. .....................89
Figure 4.15: Adjacent retinal vessels in a live pig. The left vessel is a vein and the
right is an artery. ....................................................................................................91
Figure 4.16: First aiming technique with parallel L-beams. .......................................92
Figure 4.17: Improved aiming technique (first iteration). ..........................................93
Figure 4.18: New ROx transportation and aiming system (left) and mobile
computer/electronics desk (right). Degrees of freedom in addition to those
offered by the wheels are indicated by arrows on the jack: a foot pump for raising,
a handle for lowering, a circular handle for twisting about the vertical axis, and a
knob for raising or lowering the back end of the table. .........................................94
10
Figure 5.1: Mesh frame for ROx cover, with openings for the eye piece and
actuators, and a protrusion in the back for the IR targeting laser. .........................99
Figure 5.2: Test target; image acquired in trial run...................................................100
Figure 5.3: Electronics and optics carts connected and in position to acquire retinal
images. .................................................................................................................101
Figure 5.4: Retinal vessels imaged at 514.5nm. It appears that the left vessel is a
vein and the right is an artery. The bright region is a ghost reflection from one of
the focusing lenses. ..............................................................................................102
Figure 5.5: Left to right, the medical cart, the optics cart, and the electronics cart in
the large animal lab. The test target is protruding past the eye piece, and the silver
hose is to the second fan which acts to remove hot air from within the black
cover. Trash bags are used to protect the electronics from any fluids involved in
the experiment. .....................................................................................................103
Figure 5.6: Pig eye sutchered open for imaging. ......................................................105
Figure 5.7: Dr. DeLuca performing a femoral cut-down to access the femoral vessels.106
Figure 5.8: By-hand calibration analysis compared to the off-axis intravitreal
illumination calibration line in vivo for same SaO2 values (equations shown). The
line of best fit is shown for the first attempt at analyzing each image.................110
Figure 5.9: By-hand calibration analysis compared to the off-axis intravitreal
illumination calibration line in vivo for same SaO2 values (equations shown) with
one wavelength omitted. The line of best fit is shown for the first attempt at
analyzing each image. The 496.5 nm point was omitted from all fits except for
the point corresponding to SaO2 = 42%, where 476.5 nm was the clear outlier. .110
Figure 5.10: Example of set of averaged vessel profiles displayed in order to allow
the user to select the region of the profile that is actually the vessel. The gray box
represents a typical user selection. “Std” indicates standard deviation of the
vessel values used; “Uncertainty” in this case is the standard error of the mean
divided by the average vessel value. The “Highest Std” and “Highest
Uncertainty” are pulled from the wavelength with the greatest standard deviation.111
Figure 5.11: First attempts at automated analysis of the vessel OD. The top graph
shows one user’s first attempt at analyzing the full set of images with the
automated process (optimized ROI). The bottom graph shows the results with an
OD spectrum whose parabolic fit had an R2 value better than 0.99; this includes
results from 4 separate analysis attempts for each image by the same user.
Uncertainties are included as error bars, ranging from 0.6-3.6% SaO2. ..............114
Figure 5.12: Comparison of calibration results using two different metrics for vessel
curve fitting: the R2 value of the parabolic fit to the OD spectrum (left) and the R 2
value of the vessel curve fit (right). Both plots use only the first attempt at vessel
analysis. ................................................................................................................116
11
Figure 5.13: Calibration lines from combining results from the two analysis metrics.
The top graph shows all data points, where error bars are the SEMλ. The bottom
plot uses only points for which the SEMλ is less than 1 nm, corresponding to an
uncertainty of about 3% SaO2. .............................................................................117
Figure 5.14: Example of vessel profiles corresponding to the vessel selection in the
bottom right image and its calculated width. The image is the sum of all 5
wavelength images co-aligned. The dark region at the top of the image is caused
by vignetting. The region selected is an artery, but a vein can be seen in the
bottom left corner of the image. ...........................................................................120
Figure 5.15: ROx in the pig lab after several improvements, including the jackmount
for the optics breadboard. Optics are uncovered here for pre-experiment
alignment and tests; the same black cover was put in place before imaging the pig
eye. .......................................................................................................................122
Figure 5.16: Paralysis of the eye via injection of lidocaine into the sub-Tenon region
of the pig eye. Eye clamps are used to hold the eye open. .................................124
Figure 5.17: Comparison between images collected by the frame grabber board (left)
and via the PicoScope (right). These images are the sums of all 5 wavelengths coaligned. There are reflections off of the focusing lenses in both images, but the
resolution difference is clear. Also note that ringing (vertical lines on the right
side of the frame grabber image) is not present in the PicoScope image. The
image on the left is from the first calibration experiment, and the image on the
right is from the second........................................................................................125
Figure 5.18: Comparison of glint from vessel images acquired with a 2 mm pinhole
bisected by a 200 μm wire (left) and by a 400 μm wire (right). ..........................126
Figure 5.19: Calibration line from combining results from the average analysis
metrics, where error bars are the SEMλ. Note the scale is different than that of
Figure 5.13. The error bars range from 0.99-4.27 nm, corresponding to an
uncertainty of about 3-13% SaO2. ........................................................................128
Figure 6.1: Checking optical quality of eye as well as becoming familiar with
location of retinal vessels for aiming purposes. ...................................................137
Figure 6.2: Anesthetized pig positioned for imaging: eye is sutchered open with
saline drip line (not used in later experiments). Note: the drip line is not
connected to saline bag yet. .................................................................................138
Figure 6.3: Left, Dr. DeLuca has identified the cecum. Right shows cecal ligation.139
Figure 6.4: Pig sewn up after cecal ligation and fecal slurry insult and placement of
Foley catheter. ......................................................................................................140
Figure 6.5: Entire experimental setup: swine after septic insult and intubation and
ROx in position to acquire retinal images............................................................143
Figure 6.6: The operator’s view of the ROx. The computer CPU, oscilloscope,
actuator driver, and slow scan signal generator are visible on the electronics cart.
12
The back of the optics cover can be seen, showing the open flap by the Ar ++ laser
(to the left) and the room allowed for the IR laser position. ................................144
Figure 6.7: Progression of temperature and heart rate during the progression of
sepsis, where time 0:00 represents closure of the abdomen after septic insult. ...145
Figure 6.8: Progression over time of Mean Arterial Pressure (top left), Central
Venous Pressure (top right), Cardiac Output (middle left), Systemic Vascular
Resistance (middle right), Stroke Volume (bottom left) and mixed venous Oxygen
Saturation (bottom right). Time 0:00 represents closure of the abdomen after
septic insult. .........................................................................................................146
Figure 6.9: MAP and CVP plotted with marks (+) indicating times at which pressors
and fluids were administered. The position of the marks on the y-axis is
determined by experiment; they have no actual numerical value. .......................148
Figure 6.10: Progression over time of SaO2 as measured via the CO-Ox for all
experiments with pertinent data recorded. Changes to the %O2 inspired are
marked by +/-, depending on whether the change was an increase or decrease.
Again their positions on the y-axis are determined by experiment. ....................150
Figure 6.11: Arterial measurements made on Pig 7 via the ROx (wavelength
corresponding to the minimum retinal arterial OD’s, corresponding to SraO2), COOx (%SaO2), and Pulse Ox (%SpO2).
The error bars associated with the
wavelengths are the SEM between fit metrics as discussed in Chapter 8. Pig 7
was the only sepsis experiment for which vessel identification was clear and a
significant number of arteries were analyzed. .....................................................151
Figure 6.12: Progression over time of central venous SO2 (top) as measured via the
CO-Ox. Changes to the %O2 inspired are marked by +/-, depending on whether
the change was an increase or decrease. Again their positions on the y-axis are
determined by experiment....................................................................................153
Figure 6.13: Top left: wavelength corresponding to the minimum retinal venous
OD’s (corresponding to SraO2) for all experiments. All other plots venous
measurements made via the ROx (wavelength corresponding to the minimum
retinal arterial OD’s, corresponding to SrvO2), CO-Ox (%ScvO2), and Swan
(%SvO2) for each individual experiment. The error bars associated with the
wavelengths are the SEM between fit metrics as discussed in Chapter 8. ..........154
Figure 6.14: Comparison of oxygen extraction ratios ascertained from femoral
vessels (right, from Pig 5, using CO-Ox values) and from retinal vessels (left,
from Pig 7, using ROx images and Chapter 8 analysis methods). .......................158
Figure 6.15: Retinal image produced via frame grabber. The raw interlaced image
was the only image available to the ROx operator when the image was first
acquired, and it was roughly a third the size of the computer screen. This is an
example from a pig eye in vivo (Pig 5). Vessel profiles (blue lines) of the region
selected in the bottom right image (the sum of all 5images co-aligned). Compare
13
to the images on the left. It is impossible to calculate the OD values from any of
the 3 wavelengths in the left column because their minimum relative intensity
value appears to be less than zero. Note that the red lines are the vessel fits, and
they are not good fits in any wavelength for in this case, probably due in part to
the presence of the glint and the lower resolution due to the orientation of the
selection. Even with the frame grabber, resolution is ~6x better in the rows than
in the columns. .....................................................................................................162
Figure 6.16: Example of images the user sees within 20 seconds of snapping an
image. Note that the image on the left normally takes up the left half of the
computer screen, and the image on the right scales similarly. The raw interlace
image (right) is still shown, but the separated images are also shown (left). It is
apparent from the separated images that there is a blur in the vessel about 2/3 of
the way down the image. It is consistent with a corneal artifact seen with an
ophthalmoscope, but it is not obvious in the raw interlaced image. This example
was taken from a pig eye in vivo (Pig 7). .............................................................164
Figure 6.17: Separated images from all 5 wavelengths at full resolution, after
binning. Note the improvement to the image quality (primarily due to noise
reduction) as compared to Figure 6.16. The improvement is not, however,
necessary for a quick qualitative analysis of the quality of the data. ...................165
Figure 7.1: Head rest for human patient, as seen from the point of view of the patient.
The horizontal white bars are the chin and forehead rests, and the height of the
chin rest can be adjusted via the silver knob on the right bar of the headrest. The
entire assembly can be raised and lowered electronically. ..................................172
Figure 7.2: Spectral sensitivity of the primary PMT used in the ROx. The PMT is a
Hamamatsu H10722-20 model. ...........................................................................177
Figure 7.3: Retinal images collected from the healthy volunteer using the ROx before
the PicoScope was implemented. Note the plot in the bottom right: it corresponds
to data from the vessel in the top right, showing that with careful analysis of an
image by hand, a reasonable OD spectrum can be determined and fit, producing a
minimum that corresponds reasonably well to the SvO2 of a healthy human. .....180
Figure 7.4: Image of healthy human retina acquired March 26, 2014. Left: image
separated into its 5 wavelengths (and scaled blank, which is completely dark).
Note the difference in contrast in the 457.9 nm image. Right: all 5 images coaligned and summed. The rectangle indicates the region of interest for the
profiles in Figure 7.5. ...........................................................................................181
Figure 7.5: Vessel profiles in all 5 wavelengths and the OD spectrum resulting from
the automated fits. ................................................................................................182
Figure 7.6: Illustration of the light paths entering the field of view of the PMT. The
blue lines indicate the return path from the retina (or other object). The solid red
lines show the paths of the beam incident on the pellicle, and the dashed red lines
14
roughly indicate the cone of scattered light from the pellicle. The shaded region
shows the cone of scattered light that reaches the PMT unless it is blocked by the
physical block. .....................................................................................................184
Figure 7.7: Normalized spectra measured by dividing ND-filtered Spectralon by
unfiltered Spectralon. Measurements were made before and after the block
realignment. .........................................................................................................185
Figure 7.8: Retinal image (summed over all wavelengths; the blank edges of each
wavelength were not illuminated evenly, resulting in darkened edges in the
summed image) in healthy human. Region of interest designated as rectangle on
image at the top right. Note the visibility of the optic disk in the top right of the
image. The bright spot in the center left is a reflection from the one of the
focusing lenses. An example of the automated analysis of a horizontal vessel is
shown to the left, with resulting OD plot shown below. Note the resolution of the
profiles. ................................................................................................................186
Figure 7.9: Image of retinal vasculature in a healthy human eye. Note the region of
interest selected contains an artery-vein pair, and consider that when using a
rectangular region of interest, it would be difficult to select a significant region
with only one of the vessels. However, an addition to the automated program
allows the user to select only one of the vessels. An example of the resulting
profile is shown on the right, where the user has selected the vein (darker vessel
on the left). At this point in the experiment, the SSM signal is not perfectly
synchronized with the data acquisition, so there is some image doubling at the
bottom of the image. This image is the sum of all 5 wavelengths, but the edge
blanks were not evenly aligned, resulting in the darkened edges in the summed
image. ...................................................................................................................187
Figure 7.10: Images of a human retina near the optic nerve head (summed over all 5
wavelengths, with darkened edges due to poor alignment of the edge blanking).
The rectangles indicate the selected region of interest (ROI). Note that a veinartery pair is present in both ROIs. The white spots are reflections from the
focusing lenses, and the SSM timing signal is not perfectly in sync with the
software, resulting in slight doubling at the bottom of the image. There is
vignetting in the image on the right, but both of these images are in decent focus.
In fact, the glint is visible in the artery (bottom of the pair) in the image on the
left. .......................................................................................................................188
Figure 7.11: De-interlaced images from human retina collected with the 7.9 mm
pinhole. The artery is the lower vessel, and the vein is the upper vessel,
evidenced by the slight difference in color (veins are darker in the blue-green,
especially at 514.5 nm). The white spots are reflections from the focusing lenses,
and the white vertical line is dust on the SSM (which moves vertically with
respect to the image). ...........................................................................................190
15
Figure 7.12: Vessel profiles for the artery (left) and vein (right). Due to the shallow
nature of the profiles, the reliability of the vessel fits is questionable. However,
the parabolic fits to the OD spectra have minima at 490.3941 nm and 484.5519
nm for the artery and vein, respectively, indicating an approximate difference in
SO2 of 18%. .........................................................................................................191
Figure 7.13: Image and corresponding vessel profiles (the selected region is boxed
off) from a healthy human eye. The resulting OD spectrum is on the bottom right.
The image is the sum of all 5 blue-green wavelengths. The IR targeting image
was in best focus when this image was acquired. The minimum value of the
parabolic fit to the OD spectrum is 481.6067 nm. ...............................................193
Figure 7.14: Image and corresponding vessel profiles (the selected region is boxed
off) from a healthy human eye. The resulting OD spectrum is on the bottom right.
The image is the sum of all 5 blue-green wavelengths. The IR targeting image
was purposefully defocused when this image was acquired. The minimum value
of the parabolic fit to the OD spectrum is 490.7706 nm. .....................................194
Figure 7.15: Image from a healthy human eye (sum of all 5 blue-green wavelengths).
This image was a serendipitous image acquisition because of the quality of focus,
but it was not in best focus in the IR and not readily repeatable. ........................196
Figure 7.16: Images of human retinal vessels (left) and the corresponding vessel
profiles (right). The profiles are those of the darker vessel, the selection for which
is shown in the bottom right image (all 5 wavelengths summed). The vessel
profile fits (in red) correspond to the OD spectral plot at the bottom right, with a
spectral minimum at 476.0564 nm.......................................................................198
Figure 8.1: CppSnap GUI program used for image acquisition. This is a snap shot
example of “Streaming” mode, using the IR targeting laser to image 2 retinal
vessels in a live swine. .........................................................................................202
Figure 8.2: Images separated by wavelength. Systematic noise is sometimes present
in 1 or more wavelengths (see wavelength 3 here) and must be filtered out. The
prompt allows the user to select which wavelength (if any) to filter. ..................204
Figure 8.3: If a wavelength has been selected for filtering, a couple of rows are
shown in frequency space (symmetrical about 0). The user is prompted to draw a
box around one set of unwanted frequencies. ......................................................205
Figure 8.4: A user-selected region is shown; ideally the amplitude would fall off with
distance from 0. The box has been drawn around an anomalous peak. ..............206
Figure 8.5: The 488.0 nm image before filtering (left) and after filtering (right). The
horizontal flip has no effect on the rest of the analysis........................................207
Figure 8.6: Skewed images vs. properly aligned images. All four images come from
the same set of data. The left images illustrate the skewing that occurs due to
total data grab that is not a perfect multiple of the sub-sampling rate. The images
on the right show the alignment with the automatic correction included in the
16
image preparation. Note the minor imperfection of the alignment on the right
does not affect the ability to further automatically analyze the data. ..................213
Figure 8.7: Profile of a segment of the reference signal. The blue profile represents
the raw profile, and the red is an overlay of the binary edge mask applied to the
raw profile to show the regions used for alignment between the reference and
primary signals. This segment is the first 60,000 pixels after any partial lines
have been removed from the beginning of the acquisition. .................................215
Figure 8.8: Plots displaying shift of the reference data with respect to the primary
data vs. the result of Equation 47. The plot on the right is the vertex of the plot on
the left with an increase in resolution. The noise in the right profile is systematic,
but its high frequency makes it insignificant for this analysis process. ...............216
Figure 8.9: Sum Image, which is a stack of all 5 images co-aligned. The box shows
the user-defined region of interest. Note that the region has no rows void of the
vessel, and there are no reflections or other vessels in the region. It also avoids
the edge regions where the inter-wavelength alignment is not as good. .............218
Figure 8.10: Left: image of the wlsem matrix used to determine the spectral
neutrality of each pixel. The value of each pixel is the SEM of all 5 wavelengths
at that pixel. Right: image of the subsequent neutrality mask applied to the image.
The only pixels with non-zero values correspond to pixels with an SEM below the
defined threshold (half of the mean value of the wlsem matrix in this case). .....220
Figure 8.11: Vessel straightened using the line drawn by the user (top left). Note the
slight bends in the vessel at the top and bottom. The image in the top left is a
stack of all 5 images before straightening; it shows the co-alignment of the
images. .................................................................................................................221
Figure 8.12: The same vessel straightened using correlation to the average profiles
from the vessel in Figure 8.11. Note the straightened top and bottom of the
vessel. ...................................................................................................................222
Figure 8.13: Process of division-straightening a vessel. This figure corresponds to
the 514.5 nm images. A: the image of the vessel division before any
straightening. B: binary mask corresponding to the divided image, including only
quotient values above the defined threshold. C: binary mask shifted so that the
right edge is aligned. D: image of vessel shifted analogous to the binary mask.224
Figure 8.14: The same vessel straightened using the division method in all 5
wavelengths..........................................................................................................225
Figure 8.15: Regions of interest selected by user for correlation (left) and for taking
the vertical average for the vessel profile (right). The images are different. Note
the user selects only the vessel region with very little fundus when selecting for
the correlation. The user selects the largest uniform region for the average
profile. ..................................................................................................................226
17
Figure 8.16: Example of edge detection method. The actual vessel profile is a solid
black line. The dotted blue and red lines show the profile division in the forward
and reverse directions, respectively. The minima of those divided lines are used
to define the sections of the profile used for fitting the vessel (green dashed line)
and the fundus (cyan dashed line). The value of w is varied in order to optimize
the regions used for the fits. Note that while it shares an initial value with the
vessel width used in the vessel fit equation, w is allowed to vary separately from
the vessel diameter. ..............................................................................................227
Figure 8.17: Example of GUI that allows users fit a curve to the vessel intensity
profile by hand. An actual profile is shown in black on the right, corresponding to
the line drawn on the retinal image shown on the left. The red curve is the
theoretical illumination spot, the blue is the theoretical vessel intensity profile,
and green is the convolved theoretical profile, which best fits the actual intensity
profile at 514.5 nm. ..............................................................................................231
Figure 8.18: Vessel profiles for all 5 wavelengths (distinguished by line color),
before and after division by the respective linear components. The dotted black
lines indicate the fundus regions to which lines were fit. The lines of best fit for
each wavelength are shown on the left as dashed lines. The tilt-adjusted profiles
on the right are simply the raw vessel profiles divided by the respective dashed
lines. .....................................................................................................................233
Figure 8.19: Example of a vessel intensity profile that is asymmetrical, probably due
to a variation in spot size across the image. .........................................................234
Figure 8.20: Effects of linearly varying illumination spot size across the vessel
profile. Left: standard spot profile (blue) compared to spot profiles whose widths
vary linearly from left to right by a factor of vs. Right: the corresponding
convolved vessel profiles. ....................................................................................235
Figure 8.21: Fits to a vessel profile using the majority of the profile (left) vs. a
smaller segment of the profile (right). The region right of the dashed black line is
used for the vessel. The dashed cyan line indicates the fundus region; in the case
of the left set of graphs, the left and right endpoints are used to determine any tilt.
Note the difference in fits, minimum wavelengths, and the R2 values of
subsequent OD spectra, even though the vessel profiles are the same. ...............239
Figure 8.22: Vessel fits resulting from different metrics. Left: R2 of parabolic fit to
OD spectrum is used as metric. Right: R2 values of vessel fits are used as metric.
Note the differences in the resulting parabolic fits: 0.998 vs. 0.909. “Min.
Wavelength” is the wavelength corresponding to the minimum OD. .................242
Figure 8.23: Vessel fits and corresponding OD spectral fits. Left: R2 value of
parabolic fit to OD spectrum is used as metric. Right: R2 values of vessel profile
curve first are used as metric. “Min” is the wavelength (in nm) corresponding to
the minimum OD. ................................................................................................243
18
Figure 9.1: Ray trace plot (y vs. ȳ and x vs. x ) for the current ROx configuration
from pellicle to eye (forward illumination path)..................................................252
Figure 9.2: Diagram of an off-axis paraboloid (OAP) with respect to the paraboloid
on which it is based (the “parent” paraboloid). This shows the distinction
between the focal length of the paraboloid (f) and the back focal distance (BFD).
The blue lines represent the incident and reflected optical axes. The red lines are
rays traced from infinity. The focal lengths of the complete “parent” paraboloid
and of the OAP are shown, as well as the angle θ used in Equations 35 and 57 and
θOA used in defining the OAP. Note that, corresponding to the convention of
these designs, this diagram is in the y-z plane, where the optical path propagates
in the z-direction. .................................................................................................257
Figure 9.3: Y y-bar diagram showing the optical elements plotted sequentially in the
forward direction. The chief and marginal ray heights in both the x- and y-planes
are plotted. The FSM was used as the stop for this plot. The working distance
(between OAP 4 and the eye) is 53.0 mm, with the Lagrange invariant equal to
0.0149 (this is equivalent to that of the current system). .....................................261
Figure 9.4: The spot size and shape for all 5 wavelengths at the corners of a 500 x
500 μm image (Zooms 2 and 4) and at the center of the image (Zoom 3) with no
focusing lenses present. Units are in mm............................................................265
Figure 9.5: Wavefront aberration plot: optical path difference in waves for all 5
wavelengths for the system without focusing lenses. ..........................................266
Figure 9.6: Layout of the miniaturized ROx in the Y-Z plane. The ray coloring is
such that blue corresponds to on-axis, and red and green correspond to scan
angles that give ray heights of ±250 μm at the retina. The FSM and OAPs are
drawn to scale to show the space they take up in the system, illustrating the need
for the beam to come in from the X-Z plane. ......................................................268
Figure 9.7: 3D view of the forward path of the system. The optical elements are
shown as frames, and the OAPs are represented in their full parabolic shape. The
ray coloring is such that blue corresponds to on-axis, and red and green
correspond to scan angles that give ray heights of ±250 μm at the retina. ..........269
Figure 9.8: Y y-bar diagram showing the optical elements plotted sequentially in the
forward direction. The chief and marginal ray heights in both the x- and y-planes
are plotted. The FSM was used as the stop for this plot. The working distance
(between L 4 and the eye) is 47.2 mm. ................................................................270
Figure 9.9: The spot size and shape for all 5 wavelengths at the corners of a 500 x
500 μm image (Zooms 2 and 4) and at the center of the image (Zoom 3) with
focusing lenses present. Units are in mm............................................................271
Figure 9.10: Wavefront aberration plot: optical path difference in waves for all 5
wavelengths for the system with focusing lenses ................................................272
19
Figure 9.11: The schematic of the return path, with blue rays representing an
illuminated point on the retina, and the green rays representing scattered light 250
μm up and over from the illuminated point that returns through the system.
Notice that the green rays pass through the FSWF, while the blue rays are
blocked. Again, the OAPs, FSM, and PMT are drawn to scale to indicate the
space they will take up in the system. ..................................................................275
Figure 9.12: Spot diagrams of the beam returning from the eye in the current system
(left) and the miniaturized system (right). Notice that the spots in the
miniaturized system maintain more chromatic uniformity than their current
counterparts due to the use of achromatic focusing lenses. The spot sizes are
listed quantitatively to the left of each spot. ........................................................276
Figure 10.1: Wire-on-paper target. Segments of 28 gauge wire are taped in three
orientations to a resolution chart printed with 1 line pair per mm. ......................283
Figure 10.2: Two sets of images acquired by the ROx of the wire-on-paper target.
The only difference between the right set of images and the left set is the lateral
position of the FSWF, demonstrating the effectiveness of the filter in blocking the
glint from the wires. The FSWF used here has a pinhole diameter of 2.38 mm
bisected by a wire 550 μm in diameter. The image outlined in red (bottom center)
is a photograph of the same region of the target imaged by the ROx. The
distortion introduced by scanning is clearly evident in this comparison. ............285
Figure 10.3: Images comparing all six test filters, with the reflection passing through
the filter (left) and best blocked by the filter (right). Diameters are listed to the
left in units of mm. ...............................................................................................286
Figure 10.4: Comparison between 400 and 550 μm bisecting wires in test FSWFs.
Pinhole diameters are 2.38 mm in both cases. The left image pair and
corresponding profiles compare the glint removal on the diagonal vessel. The
right image pair and profiles compare glint removal from similar sections of the
horizontal wire in the image. ...............................................................................287
Figure 10.5: Two examples of retinal vessels in enucleated pig eyes. There is no
vessel glint present in these images, further exemplified by the vessel profiles (in
blue) to the right produced from the selected region in the top left image. The red
line fitting the vessel profiles is used to produce a high-quality parabolic OD
spectrum. ..............................................................................................................288
Figure 10.6: Proposed ROx design with plane folding and OAP focusing mirrors
replacing the focusing lenses. FM1 and FOAP 1 (in the dashed box) can be
translated forward and back, indicated by the dashed arrow. Note that an afocal
system with a perfect lens is used to model the eye instead of the Arizona Eye
Model. ..................................................................................................................290
20
TABLE OF TABLES
Table 6.1: Mean times and ranges (in minutes) from perforation of the cecum to
endpoints in the sepsis model (left column). SBP is Systolic Blood
Pressure, and SvO2 is ScvO2 as measured by the Swan, specifically. ...................147
Table 8.1: Variable parameters used in the automated data analysis (in units of
pixels, except for the percentage of spot variation). ............................................236
Table 9.1: On-axis beam path data for locating limiting apertures and stops.
The iris is the aperture stop for the entire imaging system. It is also the exit
pupil, and the fast-scan mirror is the entrance pupil. ...........................................254
Table 9.2: Scanning illumination path data for finding stops. The chief ray of
the illumination beam path (no scanning) is simply a scaled version of the
marginal ray of the scanning beam path, so it is used here. For the current
configuration, the first parabolic mirror is the aperture stop. ..............................255
Table 9.3: Zernike polynomial coefficients of maximum/minimum scan angles
for a 500 x 500 μm image (Zooms 2 and 4) and of the beam on axis with no
focusing lenses present. Notice that there is less than a wave of any
aberration. ............................................................................................................265
Table 9.4: Zernike polynomial coefficients of maximum/minimum scan angles
for a 500 x 500 μm image (Zooms 2 and 4) and of the beam on axis with no
focusing lenses present. Notice that aberrations are comparable with these
lenses in place; ±45° astigmatism is the largest at about half a wave..................273
Table 9.5: The Code V lens data for the miniaturized system with the focusing
lenses in place and an ideal lens used in place of the eye (with about the
same focal length in air). This is an afocal system. The stop is set to the
entrance point pinhole. .........................................................................................273
21
ABSTRACT
The ROx is a retinal oximeter under development with the purpose of noninvasively and accurately measuring oxygen saturation (SO2) in vivo. It is novel in that it
utilizes the blue-green oximetry technique with on-axis illumination. ROx calibration
tests were performed by inducing hypoxia in live anesthetized swine and comparing ROx
measurements to SO2 values measured by a CO-Oximeter. Calibration was not achieved
to the precision required for clinical use, but limiting factors were identified and
improved.
The ROx was used in a set of sepsis experiments on live pigs with the intention of
tracking retinal SO2 during the development of sepsis.
Though conclusions are
qualitative due to insufficient calibration of the device, retinal venous SO2 is shown to
trend generally with central venous SO2 as sepsis develops. The novel sepsis model
developed in these experiments is also described. The method of cecal ligation and
perforation with additional soiling of the abdomen consistently produced controllable
severe sepsis/septic shock in a matter of hours. In addition, the ROx was used to collect
retinal images from a healthy human volunteer. These experiments served as a bench test
for several of the additions/modifications made to the ROx. This set of experiments
specifically served to illuminate problems with various light paths and image acquisition.
The analysis procedure for the ROx is under development, particularly
automating the process for consistency, accuracy, and time efficiency. The current stage
of automation is explained, including data acquisition processes and the automated vessel
fit routine.
Suggestions for the next generation of device minimization are also
described.
22
Chapter 1:
INTRODUCTION
A. Introduction
This dissertation documents the further development of retinal oximetry utilizing
the blue-green oximetry (BGO) technique, including emphasis on the ROx retinal
oximeter itself. The end goal of the ROx is to measure the oxygen saturation of retinal
vessels accurately enough to be clinically useful, and to use these measurements to
diagnose and direct therapy in a clinical setting. The ROx works by scanning the eye
with low-power laser light to image the retina with five different wavelengths to
determine the optical density spectrum of the blood.
However, device calibration
experiments have not yet shown the accuracy necessary for clinical use, and data analysis
procedures are under development.
Nonetheless, the ROx-3 has been used to help
qualitatively characterize the behavior of the retinal vascular systems of swine
undergoing the progression of septic shock.
Retinal images from a healthy human
volunteer were also acquired with the ROx-3.
B. What is Retinal Oximetry?
“Oximetry” is the measure of oxygen saturation (SO2, the percentage of oxygen)
of the blood, generally by photometric means.
Clinicians can use SO2 as a vital
diagnostic in cases such as sepsis and internal bleeding, and its potential usefulness is still
being explored. By comparing SO2 in arteries to SO2 in veins, the body’s oxygen use can
be determined.
23
Retinal oximetry is a noninvasive means of measuring the SO2 in blood vessels in
the back of the eye. The common physiological concept of “blue” veins and “red”
arteries is actually the basic principle utilized by retinal oximetry. Correlation between
the absorption spectrum (quantified color) of the blood and the amount of oxygen present
in the blood is the common basis for several different modalities, all of which illuminate
the eye and measure the return light. The uncertainty of retinal oximetry is, as of yet, too
great for it to be used in a clinical setting. A technique has previously been developed,
utilizing the minima shift of the oxyhemoglobin absorption spectrum in the blue-green
wavelength range. This is referred to as the blue-green oximetry (BGO) technique, and it
offers potential to measure SO2 with clinically useful accuracy. The ROx is a scanning
laser ophthalmoscope built to utilize BGO.
C. Description of Content
Chapter 2 describes the physiological background and implications of retinal
oximetry. The medical motivation for oximetry (and specifically retinal oximetry) is
included.
The chapter also contains a description of the anatomy of the eye, and
specifically the physiology of the retina. The optical layout of the eye is detailed as well.
Chapter 3 elucidates the BGO technique. Current clinical devices for measuring oxygen
are presented, as well as an overview of the history of retinal oximetry. The origin of
BGO, the light path analysis and spectral theory upon which it is based, and the
assumptions and implications that stem from this technique are addressed.
The
calculations for finding optical density (OD) and the oxygen saturation (SO2) are also
discussed.
24
Chapter 4 details the ROx itself.
The illumination and power delivery are
specifically addressed, including the principle of interlaced illumination. There is a
description of the acousto-optical tunable filter (AOTF) used for blue-green wavelength
switching, the infrared (IR) laser used for targeting, and the electronics integrating the
two. The power loss and safety considerations are also introduced.
There are three sets of experiments described in this work. The first is attempted
instrument calibration, imaging the retina of live anesthetized swine with the ROx, and
measuring the arterial SO2 simultaneously via CO-Oximeter (CO-Ox) while the arterial
SO2 is varied by changing the oxygen inspired by the animal.
The second set of
experiments involves inducing sepsis in a live pig and observing retinal SO2 and other
vital signs through the progression of sepsis. These pig experiments are described in
Chapters 5 and 6, respectively. The third experiment consists of using the ROx to image
the retina of a healthy human volunteer breathing air. The details of the experimental
setup and resultant data are detailed in Chapter 7.
The acquisition and analysis
procedures used for all of these experiments are described in depth in Chapter 8.
In order for this type of device to become marketable, several modifications must
be made to the ROx-3, not the least of which is system miniaturization. Chapter 9
contains an overview of previous designs, system requirements/considerations, and the
methodology behind designing a smaller system.
Chapter 10 summarizes conclusions
drawn throughout this dissertation and includes suggestions for future work.
D. Summary
This dissertation describes the current status of the ROx-3 retinal oximeter and its
application in studying sepsis and hypoxia. ROx calibration tests were performed by
25
inducing hypoxia in live swine and comparing ROx measurements to SO2 values
measured by a CO-Oximeter. The ROx was used in a set of sepsis experiments on live
swine with the intention of tracking retinal SO2 during the development of sepsis. In
addition, the ROx was used to collect retinal images from a healthy human volunteer.
The analysis procedure for the ROx is under development, particularly aimed at
automating the process for consistency, accuracy, and time efficiency. An account of
improvements made to the device is in included, as well as a description of remaining
shortcomings and related solutions and considerations.
Suggestions for device
minimization are also described. The descriptions and discussions in this dissertation
will hopefully be useful to the further development of the ROx and retinal oximetry in
general.
26
Chapter 2:
PHYSIOLOGICAL BACKGROUND
A. Introduction
In order to understand the importance of retinal oximetry, one must first
understand the physiological processes and definitions surrounding oxygen saturation.
This chapter contains relevant background on blood and oxygen transport, the central
vascular system (with emphasis on cardiac output), and finally the eye, including the
physiology and optical properties thereof. The majority of the information contained in
sections B and C is found in Boron and Boulpaep’s textbook, Medical Physiology,
Second Edition1.
The content of sections D and E was gathered primarily from
Schwiegerling’s class notes based on his Field Guide to Visual and Ophthalmic Optics2.
It is important to note that the numbers used throughout this chapter are subject to
variation, depending on individual physiology and health. They are meant primarily as
reference to the accepted normal values.
B. Blood
In ancient times, blood was considered the life of an animal or human. Science
has revealed that blood is an intricate mixture of red and white blood cells, and platelets
suspended in liquid plasma. It delivers nutrients and oxygen throughout the body to be
exchanged for waste and CO2. Modern medicine has made it possible to analyze the
health of the body by examining the blood (especially its flow, pressure, and content).
The correlations are innumerable and still emerging.
27
Plasma is the most substantial and least-dense component of blood, comprising
almost 55% of the blood volume. It is a solution (90% water) of lipids, carbohydrates,
electrolytes, and plasma proteins (including antibodies, coagulation factors such as
fibrinogen, and proteins that monitor the osmotic pressure of the circulatory system).
The bloodstream, specifically plasma, facilitates the transfer of both nutrients and waste
throughout the body. Nutrients are absorbed by the plasma and transported to tissue
throughout the body where they are exchanged for waste.
Platelets, along with coagulation proteins, are one of the primary components of
blood clots. White blood cells (leukocytes) defend the body against infection. They fight
bacteria, viruses, and parasites, and they also play roles in allergic reactions and cell
immunity. Platelets and white blood cells make up less than 10% of the blood volume of
a healthy adult human.
Red blood cells (RBCs), also known as erythrocytes, make up the fourth and most
dense component of whole blood. The concentration of RBCs in whole blood is called
hematocrit. This quantity is included in several methods for computing SO2.
Common
levels of hematocrit are 40% of the volume of RBC’s in the blood for females and 45%
for males. RBCs serve three main purposes: carry O2 from the lungs to systemic tissue,
carry CO2 from tissue to lungs, and help maintain the pH of the blood. They are biconcave cells, which is advantageous for gas diffusion. They also have no nucleus, so
they do not require oxygen for their own metabolic processes 1.
RBCs are primarily made up of hemoglobin (Hb). The heme is an iron ion (Fe2+)
pinched between 4 nitrogen and other non-metal atoms in a porphyrin ring. Hb is a
“tetramer”, meaning there are 4 hemes in a single molecule of Hb. The globin is a chain
28
of amino acids surrounding the heme, allowing O2 molecules to loosely (reversibly) bond
to the heme, but preventing the O2 molecules from irreversibly binding to the iron and
forming Fe3+. Each heme is capable of bonding to a single O2 molecule, so a single
molecule of Hb can hold 4 O2 molecules. The bonds forming the iron-porphyrin complex
are loose enough that they absorb photons in the visible range. When O2 is attached, the
bonds become slightly tighter causing the Hb molecule to absorb photons of slightly
higher frequency, leaving the reddish light to be reflected and giving oxygenated blood a
red color. This is called “oxyhemoglobin”, HbO2. Without the O2 the bonds are such
that lower-frequency photons can be absorbed, resulting in the purplish color of
deoxygenated blood. The hemoglobin is then referred to as “deoxyhemoglobin”, still
referred to as Hb1. This shift is in absorption is key to the blue-green oximetry technique
discussed in Chapter 3.
There are also several other forms hemoglobin that are not O2-carrying. For
instance, methemoglobin (metHb) contains ferric iron (Fe3+), which is already oxidized
and will not release oxygen readily. Other forms of hemoglobin bonds in RBCs include
carboxyhemoglobin (HbCO, most common after Hb and HbO2 at 1-19%, since Hb has at
least 200 time higher affinity for carbon monoxide than for O2), sulfhemoglobin (HbS),
hemoglobinfluoride (HbF), and others that may show up as a trace presence in the
absorption spectrum of blood3.
Water does not absorb any significant amout of visible light, and RBCs comprise
at least 80% of the remaining components of blood, so the absorption spectrum of blood
in the visible range is very nearly that of hemoglobin4. While some O2 is dissolved in the
plasma and RBC cytoplasm, over 98% of the O2 in the blood is carried by Hb molecules
29
in the RBCs 1. This means that like Hb, blood has slightly different absorption spectra
depending on how much O2 it is carrying (i.e. oxygen saturation). Oxygen saturation
(SO2) is defined as the fraction of the O2 capacity that is occupied by O2, or
SO2 
HbO2 
HbO2   Hb .
1
C. Oxygen Transport
Systemic tissue in the body requires O2 for survival.
pulmonary capillaries where O2 molecules bind to Hb.
Blood flows through
The heart then pumps the
oxygenated blood throughout the body via the arteries, and capillaries allow for O2
diffusion into adjacent tissue. The blood then returns for more O2 via the veins.
Because it is such a critical mechanism, blood flow in the body is very important
to quantify. One means of quantification is called cardiac output (CO), defined as the
total mean flow (F, in L/s) in the circulation (or the flow of blood delivered by the
heart)1. This is also equivalent to heart rate (HR) multiplied by stroke volume (SV),
where SV is the blood volume output of the heart in a single beat. One method of
measurement uses the Fick method (based on conservation of mass) to indirectly
determine the flow:
CO  F 
Q O2
[O2 ] B  [O2 ] A
.
2
The CO is equal to the O2 consumption rate, Q , divided by the difference in O2
concentration before and after a given point in the vascular system.
The Hb content of blood is 15 g Hb/dL blood for a healthy male, and the O2
capacity of Hb is ~1.35 mL O2/g Hb. This means that the maximum O2 content of
30
arterial blood would be ~20.3 mL O2/dL blood1. For healthy adults, arterial oxygen
saturation (SaO2) is 90-100%, while normal mixed venous oxygen saturation (SvO2) is 6080%. SO2 is [O2]/[O2]max, meaning that SO2 is directly proportional to [O2]. More
specifically, SaO2 = [O2]a/[O2]max and SvO2 = [O2]v/[O2]max. Therefore only SO2 will be
discussed for the remainder of this document with the understanding that [O2] is
effectively interchangeable.
There are several factors in the delivery of O2. More oxygen is consumed by
tissue when body temperature rises. Additionally, when the SaO2 and/or partial pressure
of O2 in the blood (pO2) is higher, O2 is more readily pulled from the bloodstream.
Increased blood flow also allows for greater O2 delivery. These values are all also
associated with elevated metabolism: as systemic tissue becomes more active, it produces
heat and CO2. The O2 extraction increases, and O2 is replaced by CO2 according to
Fick’s Law (flow is proportional to the concentration difference across a barrier). In
order to keep up with the O2 consumption, the blood flow increases.
The value of interest, especially for determining the cardiac output, is the
extraction ratio, ERO2.
ERO 2 
S a O2  S v O2
.
S a O2
All of these values are related to CO2.
3
However a phenomenon called
autoregulation prevents blood pressure and flow from being reliable measures of CO2.
The body compensates for loss O2 or blood by increasing flow and maintaining blood
pressure—especially in vital organs such as the brain and kidneys. Autoregulation does
not occur in all regions of the body. The vital organs are required for survival, so if there
31
is a shortage of blood or O2, blood flow to the extremities lessens in order to compensate
for the needs of the more important organs.
D. The Eye
The primary purpose of the eye as an organ is to allow light into the body in an
organized manner such that the light can be detected and processed by the brain for
sensing and response. As an optical system, the eye is simple: it effectively contains 2
lenses (the cornea and the crystalline lens) submersed in clear liquid (vitreous and
aqueous humors) through which light is focused onto the detector (the retina). However
in this section the workings structures of the eye will be described in relevant detail in
keeping with the scope of this work.
a. Anterior Chamber
The cornea is the clear, outer-most layer of the eye. It is clear with an index of
refraction of 1.377 and a thickness of 0.55 mm. The sclera (white of the eye) surrounds
the entire eyeball and merges into the cornea at the front of the eye, where the cornea
bulges out a bit (almost 2 mm). The cornea accounts for approximately two-thirds of the
power of the eye. Visible light is transmitted by the cornea as is infrared (IR) light.
Between the cornea and the crystalline lens, the anterior chamber is filled with a waterlike fluid called the aqueous humor.
It has an index of refraction 1.337, and its
transmission spectrum closely follows that of the cornea.
At the back of the aqueous is the pupil of the ocular system, called the iris. The
iris is the colored part of the eye (commonly described as the eye color). A primary
cause for change in pupil size is change in perceived brightness: as brightness increases,
32
the pupil diameter decreases. In bright light the iris has a diameter of about 2 mm, and it
can get as big as 8 mm in the dark.
The sensitivity curve of the human eye is wavelength-dependent: it peaks in the
green and tapers to zero near the edges of the visible range. Therefore green light
appears brighter than red or purple light of equal intensity. This also means that the pupil
will contract more for green light than for purple or red light of the same intensity. In
fact, when IR light is used, the pupil remains practically dilated at intensities that would
be considered “bright” in visible wavelengths. This phenomenon can be advantageous in
retinal oximetry, as will be explained later.
b. The Crystalline Lens
The crystalline lens is positioned about 3 mm behind the cornea. The crystalline
lens not only contributes a third of the power of the eye, but it is flexible and allows for
visual accommodation. The lens is connected to the ciliary muscle by ligaments called
zonules. When the ciliary muscle is relaxed, the crystalline lens is taught and they eye
has minimum power. As the ciliary muscle flexes, the zonules are compressed, and the
lens bulges, increasing its power. The average lens ranges in power from 15-27 diopters.
The lens has a gradient index of refraction. Starting at the center and moving
radially outward, the index ranges from 1.36 to 1.41. The lens gains outer layers of
increasing index with age. The thickening of the lens eventually results in a loss of
flexibility, limiting the ability to accommodate. The lens also significantly absorbs light
in the UV range. This causes a break down in the cells of the crystalline lens that leads
eventually to cataracts.
33
c. Posterior Chamber
Between the crystalline lens and the retina, the eye is filled with a clear, jelly-like
fluid called the vitreous humor. The vitreous has an index of refraction of 1.336, similar
to the aqueous. The vitreous enables the eye to maintain its shape (and in turn, its ocular
power) without significantly changing the transmission of light since the transmission
spectrum of the vitreous is very similar to that of the lens.
d. The Retina
The primary function of the eye is vision, i.e. detection of light. The eyeball itself
is contained by a tough, very vascular membrane called the choroid. Lining the back of
the choroid is the retina, which is comprised of all of the photoreceptors in the eye.
There are multiple layers of the retina. Photons actually bypass 3 layers and 4 types of
retinal neurons before reaching the photoreceptors and the retinal pigment epithelium
(RPE) in the outermost layers. The neurons serve as the wiring from photoreceptor to
optic nerve. They are smaller than 400 nm, so visible light is able to pass through them
effectively undisturbed. Rhodopsin is the molecule with which the light reacts in the
photoreceptors: photons change the polarity of the rhodopsin, generating an electrical
signal.
The retina is very complex, with synapses (conversion of electrical signal to
chemical signal) occurring throughout a neural sub-system within the retina itself. There
are two layers—the outer and inner layers—connected by bipolar cells. The outer layer
of the retina consists of the RPE, photoreceptors, and horizontal cells which facilitate
synaptic connections between photoreceptors as well as with the next layer. The inner
layer of the retina consists of the retinal ganglion cells, amacrine cells (analogous to the
34
horizontal cells, but for the ganglion cells), and the axons through which the signal
travels to the thalamus in the brain (the nerve fiber layer).
There are 2 types of photoreceptors: rods and cones. Rods are monochromatic,
fast-response photoreceptors that are used in the dark (they are actually bleached in
daylight conditions) and for peripheral vision. Cones, on the other hand, are sensitive to
color, but they are slower-transmitting and are not efficient in low-light conditions.
However they have a higher resolution (acuity). Cones are present throughout the macula
(the area on the retina that subtends 30° in the field of view), but they are most densely
packed in the fovea—a round, avascular region that subtends 5 degrees in the center of
the field of view. They are most densely concentrated in the foveola, which is the region
within 1 degree of the optical axis of human vision. There are actually no rods present in
the foveola. Rods and cones range in size throughout the retina from 2.5 μm to 10 μm.
Useful illustrations of the anatomy described in this section can be found in the SPIE
Field Guide to Visual and Ophthalmic Optics2.
Just below the photoreceptors is the RPE, which serves to absorb light and
prevent scattering (along with the deeper layers of the choroid and the sclera). It is
important to note that the RPE contains melanin, which has its own absorption spectrum.
The optical density of melanin peaks at around 350 nm and decays roughly exponentially
as wavelength increases, becoming nearly transparent in the red and IR regions.
Furthermore, retinal reflection varies with wavelength, but also with skin pigmentation.
However in the range from about 440 nm to about 550 nm, the reflectance is almost
constant with respect to both wavelength and skin pigmentation.
35
Figure 2.1: Image of a typical human left eye. The encircled region on the left
corresponds to the ONH, the outer encircled region on the right corresponds to the fovea,
and inner encircled region on the right corresponds to the foveola. This photograph was
taken by the fundus camera in during a routine eye exam (Permission, Douglas Parsons).
There is a region of the retina devoid of photoreceptors where the optic nerve
bundle enters the eye (also known as the blind spot or optic nerve head, ONH). The
retina is actually part of the central nervous system, since it converts light into neural
signals and processes a significant portion of the information it receives. In addition to
the nerves that transmit signal from the photoreceptors to the brain, the ONH is the
entrance point for the retinal vascular system, as well. It is 1.5 x 2 mm in diameter. The
blood vessels are most densely situated here, making it an important region for retinal
oximetry.
e. Vasculature of the Eye
The photoreceptors must be near a bloodstream for oxygen, nutrient and waste
disposal purposes. The choroidal vasculature meets that need for cells on the outer layer
of the retina. To supply blood to the inner layer, blood vessels protrude throughout the
retina (though vessels are smaller and smaller approaching the macula). These vessels
36
are actually part of the cerebral vasculature, meaning they provide oxygen to part of the
brain (in this case, the neurons that carry information about light detection). The retina
not only detects light, but it also pre-processes the subsequent electrical (neural) signal.
It is the first visuotopic map in the brain. A visuotopic map is a part of the brain that
processes spatial information gathered from the senses1. This means that as a part of the
brain, the retina is subject to autoregulation of blood flow. Consequently, retinal vessels
are useful for characterizing blood content and flow to vital organs. In addition, they are
noninvasively accessible because light has a transparent pathway to/from retinal vessels.
These are two key motivating factors of retinal oximetry.
The retinal vessels sit about 200 μm above the photoreceptors. After entering the
eye through the ONH, retinal veins and arteries branch out, generally in pairs, across the
retina. As the vessels fork and spread their diameters decrease, until they are finally
connected as capillaries. As mentioned in Section B, veins are darker than arteries in the
visible range. There is also a slight variation in size, as veins tend to be about 25% larger
than their arterial counterparts5. The largest veins are around 200 μm, and arteries are
around 150 μm.
f. Motion of the Eye
The eyeball is not a stationary organ. In order to gather input, the eyeball moves
around in its socket, controlled by the rectus muscle surrounding the eye. In addition to
the conscious motion of the eye as one looks around, there is motion called saccades,
which serves to optimize vision by constant scanning of the eye focused on a stationary
target6. Saccades occur about 2 or 3 times per second, moving 2-10 degrees in 30-50 ms
(respectively). There are also smaller movements called microsaccades. Saccades can be
37
minimized when a person stares intently at an object, but microsaccades are still present,
vibrating over several arcseconds. Both saccades and microsaccades are involuntary
motion.
E. Optical Layout of the Human Eye
In order to examine vision and retinal imaging, the light path through the eye
must be addressed. There are two main components of the eye that contribute to the
optical power of the eye—the cornea and the crystalline lens. The aqueous and vitreous
humors affect the optical path length of the eye. There are also aberrations that vary with
age and individual. This section details how the eye functions as an optical system.
First a brief primer on Gaussian optics is needed. In order for a surface to have
optical power (φ), there are two requirements: a change in index of refraction, and
curvature. Note that power is the inverse of focal length. Working under the reasonable
assumption that the surface is very nearly spherical, the power of a given surface is

1 n2  n1
,

f
R
4
where R is the radius of curvature of the surface, n1 is the index of refraction of the
incident side of the surface, and n2 is the index of the exiting side of the surface. By
definition, f is the effective focal length. Principle planes are conjugate planes along the
optical axis at which the magnification is 1. For a single surface, they both occur at the
vertex of the surface itself. However, when a thick lens (i.e. multiple surfaces) is
considered, the principle planes are offset from the surfaces. The total power of two
surfaces (i.e. a thick lens) can be calculated using
38
  1   2 
1 2 t
n
,
5
where t is the distance between, and n is the index of refraction of the medium
between the two surfaces. The positions of front and rear principle planes of the thick
lens, P and P’ respectively, are also determined with Gaussian optics.
Similarly the total power of two thick lenses can be determined with Equation 5,
using the distance between P’1 and P2 as t. If there are more than two lenses, they can be
combined two at a time. There are four sources of optical power in the eye: the anterior
and posterior cornea and the anterior and posterior crystalline lens (referred to also as
simply “the lens”).
They effectively form a pair of thick lenses that can then be
combined into a single lens. The total power of the average human eye is 59.94 diopters
(or 0.05994 mm-1), meaning the effective focal length is 16.683 mm. For the eye as a
whole, P is located 1.595 mm from the corneal vertex, and P’ is behind it at 1.908 mm
from the corneal vertex. From P and f, it can be determined that the front focal point of
the average eye is located 15.089 mm in front of the corneal vertex.
Similarly,
accounting for the index of the vitreous (f x nvitreous), the rear focal point of the average
human eye is physically located 22.289 mm from the corneal vertex.
Two other points of interest are the front and rear nodal points, N and N’. They
are defined as conjugate point at which the angular magnification is 1. In other words,
for a scanning laser ophthalmoscope, N is ideally the point about which all light should
be rotationally oriented in order to optimize the efficiency of the retinal illumination and
light collection from the eye. N is located 7.2 mm from the corneal vertex, shown in
Figure 2.2.
39
In addition to optical power, the components of the eye also introduce aberrations.
One of the most readily demonstrated is spherical aberration as light rays are further from
the optical axis. Spherical aberration is the change of focal length with respect to
aperture (iris size, in this case). Figure 2.2 shows how the spot at the retina spreads as
object height increases. The other aberration that is clearly apparent from Figure 2.2 is
coma, which is a change in magnification with aperture size and field height.
Figure 2.2: Optical layout of the eye (front nodal point labeled as N). The Arizona Eye
Model (relaxed) is shown as rendered by Code V.
There are two other aberrations of particular interest in this work. The most
common (and commonly corrected) visual aberrations are defocus and astigmatism.
Defocus occurs when the image plane is not located where the light is focused. In
40
ophthalmology, defocus is better known as myopia (near-sightedness) and hyperopia (farsightedness). In a myopic eye, light is focused in front of the retina; similarly, in a
hyperopic eye, light is focused behind the retina. These are relatively easy to correct
using simple spherical lenses. In ophthalmology, optical power is measured in diopters.
Though there is no prescription defining legal blindness, +/-10-20 diopters are the
extremes of corrective lenses. A diopter (D) is defined the inverse of the required change
in focal length in units of inverse meters.
The second commonly corrected visual aberration is astigmatism. Astigmatism
occurs when the lens is football-shaped, causing light to focus differently with respect to
orthogonal radial axes. For example, if the axes of astigmatism were vertical, the vertical
component of an image focuses to one plane (the retina), and the horizontal component
focuses to another. Correction of astigmatism is nontrivial, however, because in addition
to corrective optical power, the orientation of the astigmatism must be considered.
Other aberrations such as distortion may be present in the eye, as well. All
aberrations vary between individuals, sometimes very drastically. There are also varying
reasons for aberrations, from disease to formation of the eye at birth. Other relevant
causes of loss of vision include opacification of the cornea or lens, scattering in any of
the optical components, detached retina, and keratoconus (condition in which the cornea
has a localized bulge), to name a few.
F. Conclusions
The eye provides noninvasive optical access to autoregulated blood vessels,
clearly motivating an optical diagnostic device by which to spectroscopically measure
SO2. With a basic working knowledge of the eye and its optics, and the blood and
41
oxygen transport, the foundation is laid on which retinal oximetry can be discussed and
explored. The physiological understanding of blood and oxygen transport is required for
prediction and analysis of measurements made via retinal oximetry. The optics and
limitations of the eye must be considered in order to produce a useful retinal oximeter,
attempts at which are discussed in the next chapter.
42
Chapter 3:
HISTORY AND PRINCIPLES OF RETINAL OXIMETRY
A. Introduction
The ROx-3 is a retinal oximeter that utilizes the blue-green oximetry technique.
In order to understand the retinal oximetry and the use of the BGO technique, the history
and principles of oximetry should be described. Section B briefly describes the historical
development of oximetry, while Section C contains an overview of mathematical theory
used for oximetry. Section D describes methods and progress of retinal oximetry, with a
focus on the spectral data collection and analysis. Finally the BGO technique is detailed
in Section E.
B. History of Oximetry
Since the discovery of the role of O2 in physiology, the blood has been analyzed
for oxygen content. Initially blood gas analysis was the method of choice, and it is still
commonly used today. This involves removing blood from the body and testing it using
electrochemical methods that can only be performed once every few minutes7. Not only
is the process invasive and time consuming, but it is dangerous when the blood is drawn
from a major blood vessel (e.g. the pulmonary artery) in order to examine the central
vascular system. A vessel wall could be compromised, the patient could develop an
infection, and the amount of time the vessel is accessible is limited.
During WWII, Kramer introduced the idea of using spectroscopy to analyze blood
gas content
8, 9
. The first oximeter was developed by G. A. Millikan in 194210. It
43
measured the transmission of “red” and “green” light through the earlobe. The logarithm
of the red transmission was shown to have a linear relationship with SO2. The green light
was used primarily to measure Hb content.
Pulse oximetry is a noninvasive measure of arterial SO2 (SaO2) invented by
Aoyagi in the early 1970’s11, 12. It is now the most commonly used form of oximetry due
to its inexpensive and noninvasive nature. It can be clipped to a finger or earlobe, and
using red/IR wavelengths, it measures the absorption of pulsing arteries. However it only
measures SaO2 (not the venous SO2), and only in the extremities.
Around 1980, Zwart et al.13 developed the first clinically useful method of
analyzing SO2 spectrophotometrically. This method has been further developed and is
now utilized in the CO-Oximeter, the new gold standard for oximetry. It can make
measurements on arterial or venous blood from the central vascular system. The COOximeter (or CO-Ox) not only measures Hb and HbO2, but also HbCO and metHb.
However this method still requires blood to be drawn from the patient, which is no better
than blood gas analysis in terms of invasiveness.
C. Principles of Oximetry
As mentioned in Chapter 2, retinal vessels are part of the cerebral vasculature.
Because retinal vessels can be noninvasively probed with light, they could conceivably
provide a way to measure SO2 of a vital organ (i.e. of the central vascular system)
noninvasively. This is the motivation behind retinal oximetry.
Oximetry is a photometric method of effectively measuring the concentration of
HbO2. Therefore, in order to discuss an oximeter, relevant principles of light and lightmatter interaction should be discussed, such as attenuation via absorption and scattering.
44
a. Lambert-Beer Law
At the root of oximetry is Lambert-Beer law, which is used to relate Hb and HbO2
concentration to the transmittance of light through blood. It follows from intuition that
less light is transmitted when more particles are in the way; absorption and scattering
hinder transmission.
Lambert’s law quantifies this relationship for a linear,
homogeneous, isotropic medium (note a pure solution of Hb is homogeneous, but whole
blood is actually inhomogeneous). The probability that a single photon will be absorbed
in some small thickness, Δz, can be defined as
prob(absorption)   ( z)z .
6
N(z) is the number of photons incident on the medium of thickness Δz, where the
number of photons is a function of the distance traveled through the medium. The
number of photons, ΔN, expected to be absorbed is
N ( z)   N ( z) ( z)z ,
7
N ( z)  N ( z  z)  N ( z) .
8
where
When the limit is taken as Δz goes to zero, the equation begins to resemble the
definition of a derivative:
N ( z  z )  N ( z ) dN ( z )

  N ( z) ( z) .
z  0
z
dz
lim
9
It follows that Equation 9 can be written as a differential equation,
dN ( z )
   ( z )dz .
N ( z)
45
10
In order to apply this to a finite bulk medium of thickness l, the finite integral is
taken of both sides,
l
N (z  l)
N transmitted
ln
 ln
    ( z )dz
.
N ( z  0)
N incident
0
11
The incident surface of the medium is defined as z = 0, and any light present after
traversing the full thickness of the medium, l, is defined as having been transmitted. If
intensities are expressed in terms of photon counting and the integral is completed,
I
 e  l
,
I0
12
where I is the intensity of the transmitted light when light of initial intensity, I0, is
incident on a medium of thickness l and with an attenuation coefficient of μ. For a linear,
homogeneous, isotropic medium, this can be re-written in terms of the wavelengthdependent absorption coefficient, α(λ), such that
I  I 0 e  (  )l
13
.
Simply stated, for a linear, homogeneous, isotropic medium, intensity decays
exponentially according to the absorption coefficient (which is wavelength dependent)
and the path length of the light in the medium.
Beer’s contribution to the law is the recognition of proportionality between the
absorption coefficient and the concentration of a medium:
  c.
14
where c is the concentration and ε is the molar extinction coefficient characteristic
of the medium (blood, in the case of oximetry).
46
The combination of Equations 13 and 14, expressed in terms of the optical density
OD, results in the Lambert-Beer law. It is common in spectrometry to simplify the
transmittance to log base 10, so the final equation is
 I 
OD ( )   log(T ( ))   log    ( )  c  l .
 I0 
15
It should also be pointed out that OD values are additive. In the case of
oximetry specifically,
OD( )   Hb ( )  c Hb  l   HbO2 ( )  c HbO2  l .
16
assuming the presence of HbCO, metHb, etc. is negligible. Note that [Hb] = cHb,
and [HbO2] = cHbO2.
b. Scattering
The Lambert-Beer law, however, does not fully describe the behavior of light in
blood. Due to the presence of RBC’s, blood is not a homogeneous medium but rather a
turbid medium, meaning that scattering occurs in addition to absorption. The index of
refraction of RBC’s is very slightly but significantly different from blood plasma (1.406
and 1.345, respectively). This difference in refractive index and the larger size of the
RBC’s with respect to visible light (5-10 times larger) makes blood a weakly scattering
solution.
Twersky mathematically described the behavior of light in whole blood,
specifically multiple scattering14.
An important conclusion is that scattering and
absorption can be expressed independently of one another. That is to say, transmittance
can be written as
T ( )  Ta ( )  Ts ( )  (e  (  )cl )  [e   l  q( )(1  e   l )] .
47
17
Ta is simply the transmittance associated with absorption seen in Equation 15. Ts
is the transmittance associated with scattering, defined by Twersky15. The function q
describes the fraction of light scattered within the detection angle, φ, from the optical
axis, and β is a factor relating the refractive index, size, and position of the RBC, and the
refractive index of the surrounding plasma.
Though thorough and often cited, Twersky’s model has proven difficult to
realize experimentally. It is often simplified (usually by ignoring the scattering term) or
modified using empirical results. Using Twersky’s model, Cohen and Laing16 presented
a retinal oximetry method that attempted to account for multiple scattering in blood in
addition to absorption.
Steinke and Shepherd later tested a different model that
accounted for both absorption and scattering17, 18. They showed that scattering could be
included in the transmission equation as:
Tc ( )  Ta ( )  Ts ( )  e  d ( a s ) .
18
Similar to Equation 17, Tc is the total light transmitted, d is the effective path
length, and Σa and Σs are the bulk absorption and scattering coefficients. However it
requires calibrated measurement of Hb, which is significantly difficult in vivo.
Denninghoff, Salyer, et al. have shown that scattering has a linear effect on optical
density with respect to wavelength, using a model based on Steinke and Shepherd’s
work. They looked at both Hb and whole blood in transmission and reflection. For some
value of SO2, the resulting equation for the transmission case is,
ODT  d (aODA ( )  b  c) .
19
ODT equals the log of Tc, ODA is the optical density of pure Hb with the same
SO2, path length, and concentration as the blood sample, a is the ratio of the effective
48
path length through Hb to the effective path length through the blood vessel, and bλ + c is
the observed linear contribution of scattering on the OD in transmission19. The OD
spectrum of hemoglobin, ODA, is very nearly parabolic in shape in the blue-green spectral
range, so the form of ODT is actually parabolic as well. It was therefore concluded that in
transmission, the scattering component shifts the OD minimum by some constant value
determined by the slope of the scattering component. It was also reasonable to conclude
that in reflection, the sign of slope of the scattering component and the direction of the
spectral shift will be the opposite of those observed in transmission. This was actually
observed by the same group in the case of off-axis in vivo intravitreal illumination20.
D. Retinal Oximetry
Since the 1950’s, attempts have been made to measure retinal oxygen saturation
noninvasively. However there is not yet a definitively agreed-upon method of retinal
oximetry, largely because the uncertainties of current methods are inadequate for clinical
use. To date, retinal oximetry has been most successfully demonstrated through the
utilization of fundus cameras and scanning laser ophthalmoscopes (though OCT and
hyperspectral imaging are up-and-coming methods, as well)21-26. The most common
approach is to measure the hemoglobin absorption spectrum, which is known to shift
between oxyhemoglobin (HbO2) and deoxyhemoglobin (Hb) (i.e. with the oxygen
saturation). This requires the ability to gather spectral data, which in turn requires the
ability to illuminate with and/or detect multiple wavelengths. Such methods include
illuminating the eye with white light and filtering the returning light27,
28
; imaging
spectroscopy, which uses a grating to scan through the illumination or reflection
(absorption) spectrum29-31; illuminating the eye with specific wavelengths simultaneously
49
and filtering the return light with filters to match the individual illumination
wavelengths32-35; and multispectral scanning laser ophthalmoscopes (SLO), using a single
detector while illuminating the eye with multiple wavelengths successively36-42.
Safety standards limit the total power of light incident on the eye such that the
power of the light returning from the eye is a limiting factor in detection. The SLO
method and illumination-scanning imaging spectroscopy have the advantage of collecting
the return light without having to further divide it, optimizing the power at the detector
and minimizing the amount of time required to make a measurement. However these
methods do not allow for spectral information to be collected simultaneously, so speed
between spectral measurements is necessary. For adult humans, the heart rate at rest is
60-80 beats per minute. With each beat, blood vessels expand and contract, so data
obtained via these methods should be collected at a rate that is not affected by this
motion. Under this consideration, illumination-scanning imaging spectroscopy is not
currently fast enough. This leaves the SLO as the method of choice for this work.
In terms of analysis, the most common methodology is based on comparing the
measured vessel absorption to the Hb and HbO2 spectra, usually using at least one
isobestic wavelength and one non-isobestic wavelength. An isobestic wavelength is a
point at which the HbO2 spectrum crosses/matches the Hb spectrum. (More generally, for
2 substances that can become one another, their respective extinction coefficients are
equivalent at isobestic wavelengths.) Because OD is related to concentration (recall
Equation 15), and concentration is related to SO2 (recall Equation 1 from Chapter 2).
Using 2 wavelengths, Equation 15 and Equation 1 can be combined, giving the SO2 in
terms of OD’s (measured), and extinction coefficients (recorded in tables, e.g. Zijlstra):
50
SO2 
OD(1 )
 Hb ( 2 )   Hb (1 )
OD( 2 )
OD(1 )
 Hb ( 2 )   HbO2 ( 2 )   HbO2 (1 )   Hb (1 )
OD( 2 )

 

.
20
It is now apparent how using an isobestic wavelength can simplify the math: if λ 2
is isobestic, Eq. 20 becomes
OD(1 )
 Hb ( 2 )   Hb (1 )
OD( 2 )
SO2 
.
 HbO2 (1 )   Hb (1 )
21
However, the high error obtained with this method of analysis (specifically
uncertainty in the SO2 calibration curve) makes it unusable in a clinical setting. Many
attempts have been made to improve the error in this method: increased number of
wavelengths, improved devices, variations on wavelengths used and combinations
thereof, etc.27, 36, 43, 44. The lowest error measurements (standard error of the mean) using a
method along these lines was made by Smith44 at ~5%, which is still greater than the
clinically allowable 3%27.
Recently, methods have been developed which involve averaging large amounts
of data from retinal regions, but there is still great variability in SO2 measurements28, 35, 42.
51
Figure 3.1: Optical density spectra of hemoglobin (Hb) and oxyhemoglobin (HbO2)45.
E. Blue-Green Oximetry
In 2005, Denninghoff et al. filed for a patent describing a new way to analyze the
spectrum of a retinal vessel using at least 3 wavelengths in the spectral range from 460523 nm
46
. This is now known as the blue-green oximetry technique, or BGO. The
process entails measuring the OD of a vessel in at least 3 wavelengths and fitting a
parabola to the spectral data. In this wavelength range, both Hb and HbO2 have a
parabolic spectral shape, horizontally off-set from one another; as the SO2 increases from
0% to 100%, the result is effectively a shift in the spectral parabola, from the Hb
spectrum to the HbO2 spectrum. This shift is measured by plotting OD vs. wavelength,
fitting a parabola to the data, and finding the minimum of that parabola. Denninghoff
and Salyer19, 20 later showed that the spectral shift of the minimum linearly corresponds
with SO2. In other words, there is a direct relationship between the wavelength at which
the minimum occurs and SO2.
52
Figure 3.2: Three light paths of interest when a retinal vessel is illuminated by a
converging cone of light: (a) the glint, which reflects directly from the surface of the
vessel and conveys little information about the oxygen saturation, and (b) scattered light
that has interacted with red blood cells, either in transmission or in backscatter.
In order to make measurements to this end, there are several assumptions to
consider. First, how does the light interact with the vessel and the surrounding avascular
fundus?
Since reflection, scattering, and absorption are all possible light-matter
interactions at/in the eye, what light is actually coming back from the eye? As described
in previous work20, 47, there are 3 main light paths involving the blood vessel that are
taken into account when considering light interacting with a vessel: (i) the “glint”, or the
specular reflection off of the vessel surface (Figure 3.2a, (ii) light that is absorbed by the
blood, and (iii) light that is either backscattered by the blood cells into the collection path
or scattered and reflected off of the fundus into the collection path (Figure 3.2b). Note
53
that the glint is a problematic element, since it contains little or no spectral information;
the scattered light contains the spectral information needed for BGO measurements.
To study BGO in a setting without the glint, an in vivo experiment was devised
that invasively illuminated the retina from an angle to the optical axis20. The retina was
imaged from the path along the optic axis, allowing the collection of spectral data from
retinal vessels and fundus without the presence of the glint. Calibration tests were
performed on 4 pigs. This experiment showed a clear and clinically useful relationship
between the OD spectrum of the retinal arteries and the arterial SO2.
A more complete description of light paths for retinal imaging on-axis actually
considers backscatter and the scattering in transmission as two separate components,
especially because the contribution of scattering to the measured spectrum of blood is
related to the scattering direction, as mentioned in Section C. There are other light paths
that have been considered, such as double-pass light that passes through the vessel twice,
but they are shown to have such a small contribution to the collected light that they are
negligible20, 47-49.
Another assumption is the spectral neutrality of light reflected from the fundus 50.
Using sub-retinally placed Spectralon, Salyer et al. showed that the fundus of live swine
eyes is indeed neutral enough to use as a spectral reference. The fundus is not truly
spectrally neutral in the blue-green wavelengths, due, for example, to the presence of
capillary beds and retinal pigmentation (especially melanin)51. However, melanin has a
linear spectrum in the BGO wavelengths and the OD is two orders of magnitude less than
the OD of Hb52. A linear contribution to the parabolic spectrum will produce a shift in
the minima, but it does not affect the slope of the SO2 calibration curve—only the offset.
54
Nonetheless, this is a spectral factor that should be kept in mind in future work,
particularly on human eyes.
In keeping with the reasoning in Section D, all measurements are assumed to be
made with a SLO.
Measurements made on-axis with an SLO cannot be made
noninvasively in transmission, however, measurements made from light re-emitted from
the eye can be thought of similarly. As described by Denninghoff et al.20, 47, the following
simplified equation is used, assuming the fraction of light not scattered or absorbed is
negligible:
Rv ( ) ( )
I ( )
.


S1
Rr
I 0 ( )
22
Under this assumption, this is analogous to measurements in transmission, or
better yet, in an integrating sphere: return light that does not interact with the vessel has
effectively the same spectrum as the incident beam (I0), and light that is not absorbed is
analogous to I, detected transmitted light (or backscattered and transmitted light, in the
case of the integrating sphere). Rv is light that is returning from interaction with the
vessel (excluding the glint), and S1 is light returning directly from the avascular fundus;
these are values that are measured directly. Rr is light reflected from the retina (relatively
constant in this spectral range, as described above). Θ is the light containing Hb spectral
information. Using this reasoning, Eq. 15 can be rewritten using values that can be
directly measured on-axis:
 R ( ) 
 .
OD( )   log v
 S1 
23
It was also shown that scattering contributes the shift in the parabola, but it is
assumed to be a linear component (recall Equation 19); this is one reason why it is
55
necessary to determine a calibration line for each new generation of the measurement
device, showing the correlation between the OD spectral minimum and SO2. The slope
of the OD-SO2 relationship should be constant, but the vertical offset is not19, 20, 47.
There are several advantages to this method.
The measurements are not
dependent on the intensity itself, but the relative OD; the relative intensity difference
between the vessel and a spectrally neutral target (like Spectralon, or even the fundus) is
the quantity of interest. This means that power variability between wavelengths is not as
problematic when compared to previous methods. Note that no extinction coefficient,
light path length, nor hematocrit is necessary for these calculations, decreasing
measurement variability.
Most importantly, the BGO technique also offers greater
accuracy, as shown with in vitro experiments, as well as in vivo experiments using offaxis intravitreal illumination19, 20, 47. The accuracy of in vivo experiments using BGO with
on-axis (non-invasive) illumination is documented in Chapters 5-7 as part of the work
described herein.
The remainder of this dissertation describes the design and implementation of a
retinal oximeter (the ROx), built specifically for the purpose of furthering the BGO
technique. The device, data analysis, and experiments in which it is used are detailed, as
well as further implications of BGO and spectral data collected from the retina in this
wavelength range.
56
Chapter 4:
DEVICE DESCRIPTION
A. Introduction
Now that blue-green retinal oximetry has been defined, it is time to discuss its
implementation. We have designed and constructed a Retinal Oximeter (the ROx) for the
blue-green method. There are many considerations that must be made in order for the
ROx to be effective and safe; these are discussed in Section B. Section C is an overview
of the optics of the device. The rest of this chapter deals mainly with illumination aspects
of the ROx. Section D details the scanning system, including the concept of interlaced
images and wavelength switching. The power delivery, including power loss in the
system and laser safety considerations, is discussed in Section E. Finally, vessel targeting
and identification are described in Section F, followed by a summary in Section G.
B. Apparatus Requirements
For the experiments performed in this research, the ROx-3 must meet certain
minimal requirements in order to be useful, safe, and functional for the BGO technique.
The ROx must be portable. This includes the lasers (targeting and blue-green)
and optical system, the image acquisition electronics and computer, and for current
practical purposes, any tools or equipment needed for quick repair or adjustment (such as
an oscilloscope, a power detector, etc.). This requirement is important for the long-term
goal of using the ROx as a mobile medical device, able to move from room to room or
person to person in clinics, hospitals, emergency departments, etc. However mobility is
57
immediately imperative for the live swine experiments, including calibration of the
device and the septic pig experiments. The animal laboratory where these experiments
are performed is located in the basement of the University Medical Center, across the
street from the laboratory in which the ROx is assembled. In other words, the ROx must
be able to be transported through doorways, across asphalt and sidewalks and linoleum
tile, and up and down elevators without displacing the optics in order for these
experiments to be executed.
In order to utilize BGO, the illumination should include at least 5 wavelengths
within the range of 460-523 nm46. Because BGO is based on fitting a parabola to the OD
spectrum of blood, at least 3 points are required. However, in order to increase accuracy
of the fit, more than 3 points are needed. Denninghoff et al. have produced calibration
lines with less than 3% error using only 5 wavelengths to fit the OD spectral data47. All
wavelengths must be co-aligned in order to properly compare vessel OD with respect to
wavelength.
This avoids imaging different sections of the vessel or vignetting or
defocusing in different wavelengths.
The necessary size of the imaged region must also be taken into consideration.
Salyer and Denninghoff successfully operated under the assumption that the avascular
fundus can be treated as spectrally neutral in the context of BGO19, 20, 53, which was further
justified by their subretinal Spectralon experiment50. There are advantages to this method
of measuring vessel OD, the greatest of which is the exclusion of the spectral effects of
the vitreous, the cornea, and the lens. Those spectral factors cancel because they are not
only in the light from the vessel, but they are also in the light from the fundus. Let the
entire spectral absorption factor of the eye before the retina be called s(λ), some
58
wavelength-dependent fraction of light absorbed by the vitreous, lens, aqueous, and
cornea. When the OD is calculated, the I/I0 term becomes
I s ( )  Rv ( )

.
I0
s ( )  S 1
24
In order to use this simplification, the scan size should be at least double the size
of a retinal vessel (large vessels are 200-300 μm in diameter) in order to gather light
returning directly from the avascular fundus in addition to the light scattered by the
blood. Therefore the imaging scan size at the retina must be ~500 x 500 μm to measure
blood vessel and surrounding fundus.
The upper limit of the scan size is determined by
the angular range of the scanning mirrors and the resolution of the detector/frame
grabber.
Other considerations must be made in order to ensure that the image is consistent
through all wavelengths. For instance, because of vessel dilation and contraction with
heartbeat, the entire data set needs to be collected faster than the rate of a heartbeat (2
Hz). “Faster” more specifically means fast enough that vessel contraction/dilation and
saccades are not significantly noticeable in the data. This time limitation eliminates
methods that require a longer exposure, such as white-light illumination with the return
path chromatically filtered.
As concluded in Chapter 3, the scanning laser
ophthalmoscope should be the imaging method of choice. In order to image a retinal
vessel in 2 dimensions with a single laser spot, there must be two conjugate scanning
elements that create a raster scan, with a slow-scan component and a fast-scan
component. In order to meet the time limitation, a slow scan must be significantly faster
than 0.5 seconds.
59
Likewise, the detector and imaging component must be fast enough to record data
at this rate. The best detector options are photomultiplier tubes (PMTs) and avalanche
photodiodes (APDs) because of their analog outputs; PMTs are more sensitive, which is
desirable for the short exposure time and the limitation on illumination power of the
ROx. Regardless, the spectral sensitivity of the detector should be as uniform as possible
in the range of the wavelengths used in order to use a single gain setting for all five
wavelengths and best utilize the image bit depth in all wavelengths.
Since the ROx must be able to incorporate BGO, five appropriate wavelengths
should be included in each data set. However, they must be incorporated in a way that
effectively images the same region in all wavelengths practically simultaneously in order
to overcome the effects of saccades and vessel dilation/contraction. Therefore instead of
repeating lines or full images for each different wavelength, the ROx incorporates a
single 2-dimensional raster scan using interlaced wavelengths. This requires cycling
through the wavelengths, changing wavelength with each line (fast scan) during image
collection. The resulting raw image contains an image in each wavelength, with sample
rate in the fast axis 5 times greater than that in the slow axis. While it would be ideal to
have the same resolution in both scan directions, it is preferable to have most similarity in
timing between images of different wavelengths. This allows the ROx to collect images
at rates significantly faster than the pulse rate.
For resolution purposes, the fast scan should be fast enough that adjacent scans
are within approximately 2 μm of one another. Because of the interlacing, there are 5
wavelengths per cycle plus a reference blank (explained shortly). This means that the
fast-scan lines in each de-interlaced monochromatic image would be spaced by 12μm of
60
one another. For a 200-μm vessel aligned with the fast scan axis (i.e. with the lowest
resolution), at least 16 pixels in each wavelength contain vessel information.
The safety of the eye introduces another set of considerations governing the ROx.
There are safety limitations on the amount of power incident on the eye. Power at the
eye must be within ANSI (American National Standards Institute) standards.
ANSI
considers thermal and photoacoustic as well as photochemical safety precautions (though
photoacoustic limits only apply to pulsed systems). Because the device is a scanning
laser ophthalmoscope, a focused laser spot is scanned across the retina. This requires a
collimated beam incident on the cornea (assuming a perfect eye). For a point source (i.e.
an object at infinity which produces a plane wave—in this case, a collimated beam),
ANSI assumes no photochemical damage for exposure times (in seconds) less than t = T1,
where T1 is greater than 10 s (for the purpose of conservative calculations) and is
expressed as:
T1  1021  0.450  .
25
Note that λ is in units of μm. For exposure times, t, between 18 μs and 10 s, the
maximum permissible exposure (MPE) of a point source (e.g. a laser beam) is
MPE t  10  1.8t 0.75 103 J  cm 2 .
26
Even for the most ideally focused point (4-5 μm in diameter), small-angle forward
scattering produces a spot size of ~25μm. However, ANSI uses the dilated pupil as the
limiting aperture diameter (7 mm, as opposed to the ~1mm beam size) when determining
permissible power limited by thermal damage. Therefore in order to determine the
permissible radiant power (MPΦ, in Watts) incident on the eye, the MPE is divided by
exposure time and multiplied by the area of the limiting aperture:
61



MPt  10  1.8t 0.25 104 0.352  W,
27
which simplifies to
MPt  10  6.93t 0.25 104 W.
28
Retinal photochemical recovery time has many factors, and it is lengthy compared
to exposure time54. The standards have three rules for safe multiple exposures, and the
one that applies to this case is Rule 2. It basically states that the exposure dose is linearly
additive up to the maximum time of anticipated exposure, Tmax. The exposure from any
group of “pulses” in the “train” shall not exceed the MPE for time Tmax. Recovery time is
not specified in the standards, but Delori suggests considering a 24 hour period as the
total time over which the exposure is added55. In other words, there is no “reset” on the
exposure time within one day, so for multiple exposures, the limits instead have to be
calculated according to Tmax, the total exposure time in a 24-hour period (individual
exposure time, t, multiplied by the number of individual exposures, n). Therefore if the
total exposure is going to last more than 10 s, the dual-limit case must be considered,
where both thermal and photochemical damage must be considered (subscripts “t” and
“pc” are used to differentiate).
For 10 ≤ t ≤ T1 s:
MPE t  110 3 W  cm 2 .
29
Equation 29 holds true for all t > 10 s for λ = 500-700 nm, but for 450-500 nm, T1
determines which MPE is lower: thermal or photochemical. For T1 ≤ t ≤ 100 s,
MPE pc  C B t 1 10 2 W  cm 2 .
Otherwise, for 100 ≤ t ≤ 3x104 s,
62
30
MPE pc  C B  10 4 W  cm 2 .
31
For wavelengths (λ) between 400 and 600 nm, CB is defined thus:
C B  100.02 450 .
32
Equations 30 and 31 can then be multiplied by the area of the pupil to get the
MPΦ.
The value of λ will be the lowest BGO wavelength used by the system in
order to give the most conservative maximum. The limiting factors in data collection
speed (effectively the exposure time) are the illumination power and detector sensitivity.
Another requirement for the ROx is the ability to target the system to a blood
vessel. Due to the sensitivity of the human eye to the wavelengths used in BGO, it is
uncomfortable for the patient to have blue-green wavelengths shone for a prolonged
(albeit safe) period of time. Therefore an IR targeting laser should be included in the
ROx system. Near IR is barely detectable by the human eye, allowing for comfort of the
patient during vessel targeting. The ROx should then be able to switch to the blue-green
for data collection and analysis, but this should be no more uncomfortable than a camera
flash. The detector should be able to detect this range, and the beam must still meet
ANSI safety standards. For NIR (700nm to 1.05μm), there are two applicable limiting
equations. For 18μs ≤ t ≤ 10s,
MPE  1.8CAt 0.25 103W  cm2 ,
33
and for 10 ≤ t ≤ 3x104 s,
MPE  CA 103W  cm2 .
The value for CA is defined as
63
34
CA  100.02 700W  cm2 .
35
The considerations listed in this section are the minimum for safety and
functionality of the ROx. As the ROx is improved, more issues of convenience and
practicality will be addressed as they come up. This is further discussed in Section E.
C. Optics Overview
The light path can be summarized as an illuminating beam (a laser) that travels
through a series of optics, scans the retina, returns through the optics by approximately
the same path, and is measured by a detector. This section details this process.
a. ROx Light Path Summary
The ROx is essentially a scanning laser ophthalmoscope. An image is formed by
raster-scanning the retina. The ROx uses a “fast” scan mirror (FSM) and a “slow” scan
mirror (SSM) that scan perpendicularly to one another. The FSM scans back and forth
very rapidly (8 kHz) as the SSM completes a single swing in one direction, creating a 2D
image of the scanned area. The SSM runs at 10 Hz, so a single image can be gathered in
0.05s. Because the two mirrors cannot physically occupy the same space, they are
imaged onto one another. For targeting, the scan size at the retina is ideally about 3 x 3
mm, and for data acquisition, the scan size decreases down to ~500 x 500 μm. The scan
angles of both scan mirrors can be adjusted, but the range of the SSM is currently more
limited; this should be considered in further customization of the electronics.
BGO requires at least 5 wavelengths for an acceptable parabolic fit, so an argon
ion (Ar++) laser is the illumination source of choice. Without having to co-align 5
different lasers, the Ar++ laser beam can be filtered into 5 different wavelengths that fit
64
the BGO criteria: 514.5, 496.5, 488.0, 476.5, and 457.9 nm. The ROx also requires an IR
targeting laser. An 808 nm diode laser is introduced into the same beam path as the Ar++
beam through a cold mirror. Both beams are then cleaned up at a pinhole; the Ar++ is
removed of additional orders from the filter or modes from the laser, leaving a beam with
~1 mm diameter. The IR beam is collimated and contracted with a set of lenses and
trimmed from the oblong diode cross-section to a circular cross-section with a diameter
of ~3 mm.
The novel aspect of the ROx is that it uses BGO on axis (as opposed to off-axis
intravitreal illumination)20, 53. In other words, the ROx is a noninvasive confocal imaging
system.
This means that the illumination beam and the light containing spectral
information share the same path. In order for this to work, a beam splitter must be used.
The ROx utilizes a pellicle beam splitter after the pinhole, reflecting ~10% of the
illumination beam towards the eye.
Surface reflections and ghosting are undesirable, so mirrors are predominantly
used in the ROx. This is also one reason the pellicle is desirable over a cube or plate
beam splitter. A concern with the pellicle, however, is its angular dependence due to its
thin film properties. It requires a specific angle of reflection; the rest of the optical
arrangement and alignment is dependent on the pellicle angle. Figure 4.1 shows the
basic, unfolded layout of the optics and the beam paths, starting from the pellicle and
ending at the field stop wire filter (FSWF, used to block specular reflections) and the
primary detector (the PMT).
65
Figure 4.1: Diagram of the unfolded basic optical system of the ROx. Starting from the
pellicle, the forward beam path is shown as a solid line. The scan mirrors path shows
how the FSM, SSM, and iris are conjugate. The scanning beam paths show the behavior
of the beam at maximum scan angles. The return path and the forward path are counter
propagating until the pellicle; the long dashes show the return path after the pellicle to the
PMT. The diagram is not to scale. “OAP” denotes off-axis paraboloids.
The FSM is imaged onto the SSM through a pair of mirrors (OAP1 and 2) that
form a 4f system in order to maintain beam collimation and a lateral magnification of 1.
However due to the size of the pupil in the eye and the ANSI power limits, it is necessary
to ensure that the maximum amount of the illumination beam enters the eye, and the
maximum amount of the light containing spectral information makes it into the return
path. This means that the FSM and SSM are conjugate to the stop of the eye (i.e. the
iris). This also means the amount of scattered light collected in the return path should be
maximized.
Therefore, the transverse magnification in the return path needs to be
minimized between the retina and the final image plane in order to collect light scattered
from a sufficiently large region, containing the desired spectral information. In other
words, the ROx should be able to collect backscattered light at a single instant from a
large range of angles after interaction with the retina (consider the solid red line in Figure
66
4.1). This increases the scattered light collected in the return path. This magnification
occurs in the two confocal mirrors that image the scan mirrors onto the iris (Sph3 and
OAP4), again maintaining beam collimation.
The beam incident on the eye is therefore collimated regardless of scan angle.
The optics of the relaxed eye focus the scanning collimated beam onto the retina. The
fundus does not absorb much light relative to blood, so when the focused beam strikes the
fundus, much more light is scattered and reflected back out of the eye than when the
focused beam strikes a blood vessel in which significant absorption occurs. The scattered
light contains the spectral signal utilized in BGO.
The scattered/reflected light exits the eye and returns along the illumination path.
Note that the reflected light follows the illumination path up to the pellicle exactly. At
the pellicle, ~90% of the light (now de-scanned) is transmitted to a mirror (OAP5) that
focuses the light through the FSWF that is conjugate to the retina; the filtered light then
reaches a PMT.
The initial beam in the forward path that is split by the pellicle is ~10X greater
than the light illuminating the eye, so scattering due to the pellicle itself is not
insignificant. To address the issue of scattered light from the pellicle contaminating the
light collected from the retina, a small square block (~2 x 2 mm) suspended by a ~320
μm-diameter wire has been placed immediately after the pellicle.
When aligned
correctly, this reduces background noise in the spectrum; however, when the block is not
in place or incorrectly aligned, the transmission spectrum of the pellicle is introduced into
the return signal. 457.8 and 496.5 nm, specifically, have a very elevated signal offset, or
baseline, based on the transmission spectrum of the pellicle (Figure 4.14). While this
67
could potentially be removed by subtracting the background, the block effectively
reduces this offset to a point at which background subtraction actually introduces noise.
This issue is further addressed in Chapter 7.
The spectral information that is critical to BGO is contained in the scattered light,
specifically scattered light that has interacted with a blood vessel. However, sometimes
this scattered light from the center of the vessel is overcome by the specular reflection
(“glint”) off of the surface of the vessel, which contains little to none of the desired
spectral information about the blood. The FSWF serves to block this reflected light and
pass scattered light to the PMT, like holding up one’s hand to block a glare from the sun.
Ideally, the filter would be a floating spot that blocked the focused point of reflected
light. However, for simplicity sake, a thin wire stretched across a larger pinhole is used
instead. This only blocks a small portion of the scattered light, but it should block most
or all of the glint.
b. Details
The instrument as described here is an updated version of the ROx-3 presented by
Denninghoff et al.47. One of the unique key features of this system is that each of the
blue-green wavelengths is produced by a single Ar++ laser controlled by an acoustooptical tunable filter (AOTF) (Crystal Technology, Palo Alto, CA). The wavelengths
used for data acquisition are 457.9, 476.5, 488.0, 496.5, and 514.5 nm, since these are
emitted with the most power. 501.1 nm is also in the range, but it is difficult to produce
without bleed from the 496.5 nm mode. As with the previous design, laser light enters
the system through a pinhole filter (PH) and reflected off of the pellicle beam splitter
(10% reflected, 90% transmitted)56. The beam is raster-scanned into the eye and de68
scanned upon return using the two confocal scanning mirrors. Between the pellicle and
the retina, the illumination and return light paths are counter-propagating. Upon return,
light passes through the pellicle (now keeping 90% of the light) and is focused with a
mirror (OAP 5) onto a field stop wire filter (FSWF) by which spectrally useful light
reaches the PMT detector (Hamamatsu). The PMT is connected to a frame grabber board
(FGB) (Foresight Imaging, Chelmsford, MA). The original FSWF was a 2-mm pinhole
bisected with a 320-μm wire intended to block the specular reflection and pass scattered
light that has interacted with the blood and contains spectral information47.
The exposure safety limits make it necessary to collect as much of the return light
as possible, so the 90:10 pellicle beam splitter is a good choice. 10% of the light from the
AOTF is reflected into the eye, and 90% of the light coming from the eye is transmitted
upon return. Though specified as a 92:8 pellicle, the transmission:reflection ratio varies
substantially with wavelength as well as angle of incidence because the pellicle is a thin
film (~ 2 μm thick). This is further detailed in the Section E.
The lateral magnification of the return path should be minimized to collect
scattered light from a sufficiently large region of the retina. This means minimizing the
angular magnification in the forward path while maintaining a reasonable scan area
within the physical limitations of the scan mirrors. The distance from the rear nodal point
of the eye to the retina is 14.776 mm2, so a scan size of 3mm would need a scan angle of
about 11.6° from the rear nodal point. Considering the index of refraction of the eye, the
scan angle entering the pupil of the eye must be 8.7°. Because the forward beam path of
the ROx has a lateral magnification of 2 after both scanning mirrors (at the cornea), the
FSM and SSM must scan 17.4° to illuminate a 3 x 3 mm region of the retina. The
69
maximum scan range of the FSM is 20°, and the angular magnification is the inverse of
lateral magnification (which is 0.5 in the return path and 2 in the forward path between
the cornea and the pellicle). Therefore, due to the physical limitations of the scan
mirrors, the lateral magnification of the return path is limited to about 0.5.
Figure 4.2: Code V diagram of the return light path (compare to Figure 4.4). The
centered blue ray bundle is the reflected light path when the scan mirrors are centered.
The green bundle comes from a retinal point 250μm from the axis when the scan mirrors
are centered.
Improvements to the optical design since my involvement in this work include
changes in the geometry of the light path and the use of off-axis parabolic mirrors (OAP
1, 2, and 4), all of which reduce aberrations in the system57. The third spherical mirror
(Sph 3) was also slated to be replaced, but empirical evidence indicated that the spherical
mirror in that position actually produced significantly less astigmatism. Two focusing
lenses (L3, L4) have been added to accommodate subjects with hyper- and myopia. L4
70
was initially controlled by 2 actuators: one moved axially for focusing, and one moved
laterally for fine adjustment of beam position and to reduce any ghost reflections due to
the dual-surface nature of the lenses.
Figure 4.3: Code V diagram of the forward light path (compare to Figure 4.4). The
center blue ray bundle is the light path when the scan mirrors are centered. The outer red
and green bundles are the light paths when the scan mirrors are at - and + 3 degrees,
respectively, corresponding to a scan size of ~500 x 500 μm.
After the pig experiments, the lateral actuator was removed. L3 and L4 were
replaced with achromats, reducing spherical aberrations and cutting axial chromatic
71
aberrations in half. The focal lengths of the lenses were also doubled, allowing for a
greater depth of field when accommodating for myopic or hyperopic eyes (see Figure
4.11).
Figure 4.4: Photo of optical layout. The blue arrow into the AOTF indicates the Ar++
multispectral beam, and the green arrow out of the AOTF indicates the beam of a single
wavelength selected by the AOTF. The red arrows indicate the path of the IR targeting
beam, which then becomes collinear with the light path from the AOTF at the cold
mirror. The gold arrow indicates the return path detected by the PMT. Note that any
cardboard serves to contain stray light behind the pinhole filter which is hidden in the
photograph.
A co-aligned 808 nm IR diode laser (driver from Wavelength Technologies,
Bozeman, MT) has also been added for targeting. Because the light is barely visible, it is
72
more comfortable to human subjects and does not cause the pupil to constrict further,
allowing more light to return from the eye. Co-alignment of the IR laser, while
important, is less critical than the co-alignment of the blue-green wavelengths because it
is not used for the actual measurement. The system can be switched between targeting
mode using the IR laser and data acquisition mode using the Ar++.
For future work, it should be noted that the 2D intermediate image plane between
Sph3 and OAP4 provides a location for a reticle or other image calibration apparatus,
such as Spectralon as a spectral reference. The other intermediate image plane located
between OAP1 and 2 only affects one dimension of the image because it does not include
the SSM dimension.
One dilemma that has been particularly troubling for confocal laser scanning
methods is the central glint: the specular reflection from the surface of the vessel, which
contains little spectral information and inhibits collecting the desired spectral information
accurately. In order to solve for optical density, the measurement needed from the vessel
is the relative intensity at the center of the vessel where the OD is the greatest. However
with on-axis illumination, the specular reflection comes from the apex of the vessel
surface and overpowers the dimmer signal from light that has interacted with the blood
(Figure 4.5).
73
Figure 4.5: Comparison or retinal scans when there is no glint (left) and when the glint is
present as in increased intensity (right). These images were gathered from an enucleated
swine eye, and it was not possible to determine whether this is a vein or artery.
Salyer et al. circumvented this problem by using off-axis intravitreal illumination
instead of illuminating on axis53. The current solution for the ROx uses a field stop wire
filter (FSWF) on the return light. The filter consists of a pinhole bisected by a wire; it is
located in an image plane conjugate to the retina such that specularly reflected light off of
the retina (and off of retinal vessels) is effectively blocked by the wire. However, the
vast majority of the light that is not specularly reflected passes to the detector. Because
the useful spectral information is contained in the scattered light, this filter should
hypothetically completely remove the central glint.
To collect scattered light effectively, a region of ~1000 μm on the retina should be
collected by the system and passed through the pinhole. Similarly, the central ~200 μm
should be rejected by the bisecting wire. The sizes of these respective regions are based
on previous work by this group58. The magnification of a given point on the retina is
determined not only by the magnification in the scanning system, but also by the power
74
of the eye and the optic that focuses the de-scanned beam onto the wire. The FSWF was
originally constructed for a system with the retina-to-pinhole magnification of 2.4, where
the relay optic to the FSWF was a lens with ~30 mm focal length. The average human
eye has a focal length of 16.7 mm. The pinhole used was 2 mm in diameter with a 28
gauge wire bisecting it. The area of the retina projected through the pinhole was ~830
μm in diameter, with the central ~130 μm (vertically) blocked to stop the glint.
The system, however, was rearranged so that it could fit on a single 1.5’ x 2’
optical breadboard for transportation to and in the pig lab. The angle of the pellicle was
changed from 45° to 27°, and the final relay lens was replaced with the off-axis parabolic
mirror immediately on-hand (now labeled OAP 5). Its focal length is 50.8 mm. This
changes the magnification of the system. ROx magnification was roughly measured to be
~5 empirically, and Code V indicates a magnification of 6; for the sake of further
calculations, it is considered to be 6, due to the imprecise nature of the experiment
yielding the empirical values. This means that the image of the retina projected onto the
FSWF was larger than originally planned, so a) the system was more tightly confocal,
making the images sharper but depth of field shorter and focusing more difficult, and b)
the specular reflection spot was larger than the wire that was supposed to be blocking it.
Unfortunately, while the glint was reduced with that filter in place, a significant amount
of glint therefore continued to appear in images collected by the ROx for the original 6
pig experiments and human eye experiments. Adjusting the FSWF size would have
likely provided a solution, however this was not immediately apparent. Consequently,
another approach to the problem is to address it in image processing/analysis, which is
described in a Chapter 8.
75
D. Scanning System
The slow-scan mirror (SSM) scans horizontally (vertically in all images due to
frame grabber setting) and is controlled with a function generator. A triangle function is
used in order to produce a uniform scan at frequencies on the order of 10 Hz 37. However
a complete image can be gathered in a half-scan of the SSM, so the image acquisition
time is actually 0.05 s (see Figure 4.6). In addition to controlling the frequency and
amplitude of the SSM, the function generator signal is also converted to 20 Hz pulses to
control the frame-grabbing rate of the FGB and, after updates, trigger the PicoScope
acquisition (see Figure 4.7). The PicoScope is an analog-to-digital converter described in
detail later.
The fast-scan mirror (FSM) is a resonance oscillator with a fixed frequency of 4
kHz; it scans vertically with adjustable amplitude. With every half-oscillation, the FSM
generates an electrical pulse that not only controls data acquisition but simultaneously
controls illumination. The pulse from the FSM is split in two: one part is used to
generate the 8 kHz pulse that triggers the FGB to move to the next line within the frame
(Figure 4.10). This results in 400 pixels in the slow-scan dimension. Depending on
whether the system is in targeting or data acquisition mode, the other part is either sent to
the IR laser driver or used to generate an 8 kHz double pulse that is sent to the AOTF
control circuit. One method of determining illumination type is a hard switch between
power to the AOTF control circuit and the IR laser driver interlock. The other method,
detailed later in this chapter, is via output from a DAQ data acquisition device controlled
by the data acquisition software; this latter method was not available for the original
sepsis and calibration experiments on live swine.
76
Figure 4.6: Raw interlaced image with edge blanking. Image doubling is occurring in
the fast scan orientation (the horizontal is the fast scan and the vertical is the slow scan).
The magnified image makes the individual wavelengths and blank lines apparent.
Figure 4.7: Signals corresponding to the SSM; the SSM angle (examples shown at top) is
controlled by the voltage from the signal generator (in black).
77
Figure 4.8: The image in Figure 4.6 is separated into the 6 channels. Because each
wavelength is used after every sixth line, the raw images are not equally proportioned in
the horizontal and vertical directions. It is evident that the edges are not blanked
perfectly. The top 5 are false-color images to make the contrast more clear.
Figure 4.9: The 514.5nm image from Figure 4.6 with equal horizontal and vertical
proportions.
Because the FSM scans bi-directionally, the IR laser is triggered to turn on for
every other scan to prevent image doubling (Figure 4.10). This eliminates the need to
post-process the video mode of the FGB and therefore allows for live targeting. The ROx
is designed to create images of a 3x3 mm portion of the retina during targeting. This
78
means that the illuminated horizontal scans are 15 μm apart, which is a very acceptable
resolution for targeting purposes.
The AOTF control circuit has two major functions, the first of which is switching
wavelengths. The 5 Ar++ wavelengths are passed sequentially through the AOTF, with
each wavelength on for ~0.1 ms during the central ~80% of a single fast-scan. After
every set of 5 wavelengths, a blank scan (125 μs of the AOTF passing no light) is taken
for identification purposes (see Figure 4.8). The switching order is constant, so the blank
image allows a user or algorithm to determine which wavelength corresponds to each
separate image. However, the laser power transmitted by the AOTF is not immediately
stabilized; there is a ringing artifact that is present each time the frequency of the AOTF
crystal is changed. The ringing is also affected by the optimization of the RF frequency
through the crystal; when the crystal frequency does not best-match the desired
wavelength, the first order diffraction is weaker and the SNR is lower. For example,
Figure 4.6, Figure 4.8, and Figure 4.9 display this ringing artifact. In the raw image
(Figure 4.6), it manifests as vertical lines on the left edge of the image, perpendicular to
the fast scan direction and at the beginning of every scan, regardless of scan direction.
As the images are separated and flipped for similar vessel alignment, the ringing flips
sides (Figure 4.8); it is easier to envision the ringing with respect to each change in
wavelength in an interlaced image.
79
Figure 4.10: Signals corresponding to the FSM
The second major function of the AOTF control circuit is edge-blanking. The 8
kHz FSM pulse is manipulated into a double pulse that turns the AOTF off and on again
at the edges of every horizontal scan. The double pulses are positioned by hand using 4
potentiometers. Figure 4.10 shows the double pulses that are the input to the AOTF
80
control circuit and their part in wavelength switching and edge definition. The first
AOTF channel (Channel 0, wavelength 1) is turned on with the edge blank applied as the
FGB starts a new line. The blank ends and the eye is illuminated and imaged as the line
scan continues. Simultaneously in the very last section of the line scan, the blank is
reapplied, Channel 0 is turned off, and Channel 1 (wavelength 2) is turned on. That scan
ends and the next begins with the FSM swinging in the opposite direction.
This process repeats continually through 6 channels; Figure 4.10 illustrates a full
period plus one line. The black dashed lines in Figure 4.10 mark the time lapsed over a
single line acquisition of the FGB and are understood to be periodic. Figure 4.6 shows a
resulting raw image. Approximately 10% of each scan edge is not illuminated. Each
edge blank applies to the last 10% of one line and the first 10% of the next line. This
allows the frame grabber to better handle the background and dark current of the PMT
throughout the data acquisition by giving it a “zero” reference (even though the PMT is
too sensitive to ever truly read zero).
Note that though precision of data acquisition is slightly less critical for the
targeting laser, the edge blanking is still required for contrast in that spectral range, as
well. It was critical for image acquisition when the frame grabber was the primary means
of imaging, and the edge blanks are also useful in squaring up the line of data acquired by
the PicoScope ADC. They serve as a reference when converting from a 1-D data vector
into a 2-D image.
The AOTF is effectively a tunable Bragg diffraction grating. A piezoelectric
plate is used to create an acoustic wave in a crystal. The compression and rarefaction of
the wave produces periodic regions of lower and higher indices of refraction. This
81
periodic change in index in turn diffracts a specific wavelength of light (λ), depending on
the period of the wave (d) and the angle of incidence (θ) according to Bragg’s Law,
m  2d sin ,
36
where m is some integer. The current setup is such that after initialization by the
user, the AOTF is controlled by an entirely separate circuit from the main computer.
This means that the switching rate is not limited by the computer processor rate, nor are
there delays associated with any program runtime.
Initialization involves copying pre-written code into the command line of the
AOTF software, though commands can also be entered by hand; there is also a useful
GUI aspect of the software. The code specifies acoustic frequency of the AOTF (i.e.
filtered wavelength), power amplitude (corresponding to the power of the light that is
filtered into the system), and wavelength-switching configuration.
Once those
specifications are set, the user controls the AOTF with a manual hard-wired switch. It is
important to note that the radiant power output as controlled by the software and AOTF
daughter board does not translate directly to the radiant power output. Though the
software still serves as the interface through which the amplitude and acoustic frequency
are controlled, an external power source is required to provide up to 10V to each channel
in order to reach maximum radiant power transmitted. Initially, this power was provided
by a AA battery pack (4 batteries, 10V total). The AOTF was eventually connected to
the source powering the rest of the ROx.
82
Figure 4.11: Photo of improved optical layout. The blue arrow into the AOTF indicates
the Ar++ multispectral beam, and the green arrow out of the AOTF indicates the beam of
a single wavelength selected by the AOTF, with the dashed green line indicating the
reference path to the reference PMT. The red arrows indicate the IR targeting path,
which becomes collinear with the light bath from the AOTF at the cold mirror. The gold
arrow indicates the return path to the primary PMT. Note that the cardboard contains
stray light behind the pinhole filter.
The first 6 pig experiments were performed with the frame grabber as the primary
means of data acquisition. However, the frame grabber collects instantaneous snippets of
data without utilizing any integration time/method, and the sample size (number of pixels
grabbed per image) was not enough to satisfactorily average out the shot noise from the
83
PMT and still maintain the required data quality. A greater sample rate or means of
integration is necessary. A capacitor was placed in parallel with the PMT signal in order
to filter it, effectively smoothing the noise. This setup was used on the first 6 pigs.
However, this did not prove to reduce the noise enough. A sharper and more precise
signal filter with a cutoff frequency of 6.954 MHz, with the signal dropping to -3 dB by 5
MHz (LTC6601 on a DC1251A demo board, Linear Technology, Milpitas, CA) was
used, under the premise that the problem was high-frequency noise (possibly shot noise)
from the PMT itself. While this sample rate is not the Nyquist frequency of the frame
grabber, we wanted the limiting factor to be the frame grabber; it would also be easy to
test the frame grabber at higher sample rates.
This filter did not have any appreciable effect on the noise in the images. The
sample rate used with the frame grabber is about 6.24 MSa/s (mega samples per second).
It has the capacity to grab at 32 MSa/s, which was used to determine that the sample rate
was a significant issue. When the greater sample rate was used, the error, as calculated
by the standard error of the mean (SEM) divided by the mean, was halved by this
resolution change in a dark region of a test image; the error decreased by more in the
brighter regions. By further boxcar averaging to roughly simulate a higher resolution,
500 MSa/s was projected to provide the noise reduction required to make this a viable
medical instrument. There is no integrator for the frame grabber, so each pixel only
samples a small fraction of the data each pixel represents. By increasing the sample rate,
more data can be binned into the same number of pixels (6.24 pixels per second, each
pixel the sum of 10-100 data points); this effectively integrates the data stream so that
84
each pixel represents more of the data collected in the time frame it actually represents
(~160 ns).
The bit depth of the PicoScope is 8 bits, which is less than the Foresight frame
grabber (10 bits). However, the intention is to bin the data into an image size comparable
to the 400x780 pixel image from the frame grabber. The PicoScope will collect ~62500
pixels per line. By summing 80 pixels per bin, the range of pixel values in a 400 x 780
image becomes 0-20400. That is a significant improvement compared to the 0-1023
pixel value range in the Foresight frame grabber, so the lower bit depth of the PicoScope
is completely acceptable. The PicoScope data is actually summed into 100 pixels per bin,
giving a range of pixel values from 0-25500 and an approximate image size of 400 x 635
pixels.
Another late addition to the ROx is the use of a DAQ (Data AcQuisition device,
MCC DAQ, Measurement Computing) to automatically switch between lasers: it allows
the IR to run while targeting a vessel, but when an image is snapped, it simultaneously
turns off the IR laser (via its driver interlock) and turns on the AOTF until the image
acquisition is complete, at which time the AOTF is turned back off the IR laser is turned
back on.
A solenoid-driven shutter (requiring a 24V differential) has also been
implemented such that the shutter is open when the AOTF is on, but the shutter is closed
when the AOTF is off (and when power to the solenoid is off). This prevents any
residual light from the AOTF from passing into the system during targeting. While the
power of the residual light is negligible (picowatts), it is visible to the human whose eye
is being imaged. The shutter control is achieved using the same signal that controls
whether the AOTF is on or off. That AOTF switch signal is used to trigger another
85
analog switch, which passes -12 V or 0 V to one end of the shutter connection (while the
other shutter connection is held at +12 V constantly).
The DAQ is also the key to using both the Foresight frame grabber and the
PicoScope to collect data from the same primary PMT. For the current system, when the
PMT signal is split, the output value is also cut in half. To remedy this at no extra cost,
the DAQ is used to control an analog switch that directs the entire signal to either the
frame grabber or the PicoScope. In addition to switching lasers, a logic signal from the
DAQ sends the PMT signal to the frame grabber until an image is snapped, at which time
the PMT signal is redirected to the PicoScope (at the same time the AOTF is turned on).
When the acquisition is complete, the DAQ returns the PMT signal to the frame grabber.
This solution could have been approached in other ways, including conditioning inputs to
the frame grabber and Pico Scope. In the future, these options should be considered.
Note that while the DAQ always controls the direction of the PMT signal, there is an
override switch that allows the user to manually control which laser is on.
E. Power Delivery
All laser light incident on the eye is in compliance with ANSI standards.
According to the ANSI standards, the smallest spot can be assumed to have a 25 μm
diameter, but the limiting aperture is the dilated pupil, 7 mm in diameter. There are
effectively 400 spots per 2D scan (gathered in 0.05 s), and each spot is illuminated for
125 μs per scan. Another way to think about this is to say t is 1/400th of the total time, T,
over which the eye is scanned in a day. This means that for t = 10 seconds, T = 6.67
hours.
86
Using Equation 28 above, for a total exposure time of 10 s on a single spot on the
retina, the maximum permissible radiant power is found to be
6.93100.25 104  389.7 μW.
37
From experiments on enucleated swine eyes, it is known that 35μW is sufficient
for obtaining a useful image of the retina, so the limit is ten times greater than the power
used by the ROx.
In the event of a system failure, wherein both scan mirrors froze and the blanking
and wavelength switching failed, another set of limits would have to be considered. For
the sake of being conservative, it is also assumed that the patient continues to look into
the beam of wavelength 457.9 nm for an indefinite amount of time greater than 10
seconds. For the sake of being conservative, the maximum exposure time listed in the
standards is used: 3x104 s. Not only does thermal damage have to be accounted for, but
photochemical damage must be considered as well. Because t > T1 and 450 < λ < 500
nm, Equation 31 is used, with the inclusion of the pupil area. Plugging in 457.9 nm for λ,
the limiting case is:
MP   .35 100.02457.9  450  4  55.4 μW.
2
38
For completeness, Equation 29 is used with 514.5 nm and the pupil area:
MP   .35  103  384.8 μW.
2
39
Because the BGO method divides the total power out of the return light (see
Equation 23, Chapter 3), the absolute powers of each respective wavelength do not have
to be exactly equal at the eye. This is one of the strengths of the BGO method. However
the FGB and PMT have limited ranges over which they will detect change (i.e. between
relative zero and saturation). In order to produce the most uniform image quality for all
87
wavelengths, the respective powers reaching the PMT should be as close to equivalent as
possible.
There are multiple frequencies present in the Ar++ beam (Figure 4.12); however,
each frequency has a different power within the beam. The specifications for the JDSU
laser used in the ROx show that 457.9, 496.5, and 476.5nm comprise approximately 6%,
8%, and 10% of the beam respectively. The rest of the beam is made up of approximately
equal powers of 488.0 and 514.5 nm.
Figure 4.12: Specs provided by JDSU on the Ar++ laser used by the ROx. The graph on
the left shows how the power of each wavelength increases with current. The graph on
the right shows the percentage of which each wavelength comprises the total beam for a
given power. The current used in eye data acquisition is 7.96 A.
The goal power at the eye is 35-40 μW, considering power limitations. Therefore
rather than finding a pellicle angle that reflects all wavelengths most equally, it is better
to find an angle that reflects the greatest amount of the three low-powered wavelengths
because 488 and 514.5 nm have more power available to be transmitted through the
AOTF. The optimal angle for the 6 useful illumination wavelengths is 26°. 457.9 and
496.5 nm are the limiting factors, and they vary almost opposite one another with angle.
Figure 4.13 shows that 26° is the angle at which they are highest together (and at which
88
the beam is reflected into the system at a practical angle). 488.0 nm is very low here, but
there is so much power coming from the Ar++ laser at that wavelength that it is easily
compensated (Figure 4.12). The greatest reflection is 7.7% for 476.5 nm, and the least is
1.7% at 457.8 nm (Figure 4.14).
Figure 4.13: Data taken by my colleague Tyson Ririe, showing the intensity from the
pellicle in reflection as the angle and wavelength are varied.
Figure 4.14: Transmission and reflection spectra of the pellicle at 27°.
89
Besides the pellicle interactions, there is some loss in the system, including ~8%
at each of 7 mirrors equally in every wavelength. The JDSU laser currently in use
fluctuates significantly, both in power and in wavelength (i.e. acoustic frequency that
optimally transmits a given wavelength). However, the optimum RF value may wander
due to the AOTF crystal itself. It will be necessary for the final product to have either a
feedback loop to control the fluctuations or the laser needs to be more stable.
F. Vessel Targeting and Identification
As previously mentioned, an IR laser (808nm) has been incorporated into the
system for the purpose of targeting vessels in the human eye. While targeting with the
blue-green is safe, it is uncomfortably bright for conscious human patients. The IR beam
must be co-aligned with the beam from the AOTF.
The source is a laser diode, which is uncollimated and larger than the Ar++ beam.
It is housed with a lens for which the position can be adjusted to produce a collimated
beam. The housing is followed by a pair of lenses (L1 and L2) that serve as contractors
to bring the IR beam closer to the diameter of the beam leaving the AOTF. The pinhole
filter (PH) also serves to help re-shape the IR beam from a rectangular to a circular beam.
Triggered by signals from the FSM (see Figure 4.10), the beam illuminates the eye such
that the PMT and FGB gather an edge-blanked image in “video mode”, i.e. every other
line is illuminated, preventing image doubling in the real-time data.
The IR laser must also meet ANSI safety standards. For the targeting laser,
Equations 34 and 35 are used, plugging in λ = 808nm and multiplying by the area of the
pupil. This gives MPΦ for the IR in the conservative case of the longest calculated
exposure time:
90
MP   0.35  100.02808 700 3  55.6 mW.
2
40
The PMT has a sensitivity that is relatively spectrally flat in the blue-green, but its
sensitivity begins to drop off a bit more around 800nm. This maximum permissible
radiant power allows for brighter illumination of the eye with IR to compensate for the
PMT sensitivity curve.
The IR is not only useful for vessel targeting, but also potentially for
identification. HbO2 absorbs noticeably less NIR than Hb. This makes it possible to
distinguish between arteries and veins when there are two to compare. As can be
observed in Figure 4.15, the veins are not only darker, but also generally larger when
compared to arteries. Arteries are also tauter and have thicker walls to withstand higher
pressures than the walls of veins.
Figure 4.15: Adjacent retinal vessels in a live pig. The left vessel is a vein and the right
is an artery.
Another important aspect of targeting is the ability to aim the device. Initially the
optical breadboard was attached to two parallel L-beams (Figure 4.16). Four threaded
91
rods protruded through a plywood base and through holes in the L-beams. The beams
were secured to the rods by nuts which were used to adjust the height of the beams on the
rods. This allowed for rough vertical tilt and height control. The rolling ability of the
cart allowed for rough horizontal positioning and tilt.
Figure 4.16: First aiming technique with parallel L-beams.
The aiming was improved using an aiming base from another device with
horizontal, vertical, and axial motion as well as vertical tilt (with horizontal tilt still
controlled by the positioning of the rolling cart—see Figure 4.17). However, the optical
board is too heavy for the new aiming base, so a frame was constructed from which to
hang the L-beams from springs in order to relieve some of the weight while still allowing
for flexibility.
The weight still imposed problems as users would attempt fine
adjustments, and it was difficult to secure the position once the device was aimed
decently.
92
All of the electronics, including a desktop computer, fit on a rolling cart that
accompanies the cart holding the optics. Though it is cumbersome, the ROx is portable
and was utilized in six live pig experiments using these original mount configurations.
Figure 4.17: Improved aiming technique (first iteration).
Because the aiming capabilities of the ROx were still unsatisfactory, the mount
underwent another upgrade. Figure 4.18 shows the updated modes of transportation for
both the optics and electronics.
The optics are now mounted on an automotive
transmission jack. It offers the ability to raise/lower, twist, and vertically rotate the
system for aiming purposes (indicated by arrows in the figure), and it is much easier to
achieve fine motion control and secure the ROx in position once adequately aimed. This
jack is a significant improvement over the previous iteration; however, none of the axes
of rotation are about the optical axis and are therefore not ideal. Additionally, the wheels
93
have been replaced with air-filled tires on casters with brakes for shock absorption during
transport and stability during imaging.
The electronics are mainly housed in an old desktop tower (now on top of the
desk). The monitor now sits on the driver box for the focusing actuators, and the CPU
sits under the desk with the Ar++ laser power supply and the AOTF driver.
Figure 4.18: New ROx transportation and aiming system (left) and mobile
computer/electronics desk (right). Degrees of freedom in addition to those offered by the
wheels are indicated by arrows on the jack: a foot pump for raising, a handle for
lowering, a circular handle for twisting about the vertical axis, and a knob for raising or
lowering the back end of the table.
G. Summary
In order to be functional and useful, the ROx must meet several requirements.
Using 2 rolling carts, it is portable. The principal illumination source is an argon ion
94
laser that produces 5 wavelengths within the BGO range. Confocal scanning mirrors
control the size of the imaged region. The 4 kHz fast scan mirror is used to control
wavelength interlacing for consistency between images of wavelengths, as well as
provide acceptable resolution and acquisition speed. The image acquisition system is
sufficiently fast and has been upgraded to reduce noise, both by incorporating time
integration for each pixel and by including a reference detector by which noise from the
AOTF and laser can be removed. The scanning system also includes blanking required
for wavelength identification and image contrast (specifically for the frame grabber
software).
An IR targeting laser has been incorporated for patient comfort, vessel
identification, and user convenience because it removes image-doubling in real time.
Finally, all aspects of the illumination system meets and exceeds the ANSI safety
standards.
The ROx has been successfully built as a confocal scanning laser ophthalmoscope
capable of utilizing BGO. The device is still large and cumbersome, but the electronic
circuitry could be made more compact and efficient in the future. A further analysis of
the optics and a recommended design for a miniaturized system is outlined in Chapter 7.
A smaller device will allow for better maneuverability in aiming, hopefully to a point
similar to the EOX-238, 58 with ease of aiming and superior image quality.
95
Chapter 5:
CALIBRATION EXPERIMENTS ON LIVE SWINE
A. Introduction
Calibration is required in order to make quantitative measurements with the ROx.
The device was first tested on enucleated swine eyes and the eye of a live human
volunteer47, demonstrating the imaging capability of the device. The tests also included a
quantitative calculation of SO2 using the calibration equation found using in vivo off-axis
intravitreal illumination BGO mentioned in Chapter 320, 47. The SO2 value found for the
enucleated pig eye was within the expected range for blood vessels post mortem, and the
SO2 value found for the live human eye was within the expected range for a human vein.
However there was no means of calibration in these experiments.
This chapter deals with the first ROx calibration experiments.
Section B
describes the first trial run with a live swine. Section C details the experimental setup
and procedure of the calibration experiment. Section D contains data analysis, Section E
looks at using the analysis process to determine the vessel widths, and Section F contains
conclusions from that experiment. Section G summarizes the second attempt at a ROx
calibration experiment, and H gives overall conclusions.
B. Trial Run
As mentioned before, the ROx had been used to gather images from enucleated
swine eyes and the eye of a healthy human volunteer47. Calibration of the device,
however, requires the ability to measure, vary, and control the SO2 of the vessels being
96
imaged.
Arterial SO2 is generally very consistent throughout the body, and the
relationship between retinal venous SO2 and SvO2 in other parts of the body is not well
understood. Consequently it is ideal to calibrate the ROx with arterial saturations. This
would best be accomplished on a live patient, and dropping the SaO2 would make the
tests hazardous to humans; animal testing is therefore necessary.
The Institutional
Animal Care and Use Committee (IACUC) at the University of Arizona has a specific set
of guidelines for animal selection and treatment. The fewest number of animals must be
used, and the species should be a “low” as possible (on a scale where single-cell
organisms are the lowest, and humans are the highest) with minimal pain or trauma to the
patient.
To comply with these standards, pigs are the animal of choice59. They have
similar vasculature, eyes, and hearts. More specifically, the size of swine vessels is
comparable to adult humans, making placement of catheters comparable. The swine
heart has 4 chambers like the human heart, and it is common practice to replace human
heart valves with pig valves. The average pig eye is similar in both optical quality and
size to the adult human eye. However, they are more astigmatic and near-sighted, and the
iris is larger than the human eye. There are other animals with eyes that are better
analogs to humans (such as other primates), but as pigs are part of our food chain, they
are considered a lower species and therefore the most suitable test subjects for these
experiments under IACUC standards.
The IACUC also requires use of the fewest
animals possible. According to previous work20, data from at least four pigs is necessary
for the calibration. This chapter outlines the first of these experiments.
97
The Department of Trauma, Surgical Critical Care, and Emergency Surgery at the
University of Arizona College of Medicine planned to perform an experimental tension
pneumothorax procedure on two live swine the week before the ROx swine experiments
were scheduled to begin, and the principal investigator agreed to allow a trial run with the
ROx after their procedure was completed and before they euthanized the animals. The
goals included testing the portability of the ROx, practicing breakdown of the device into
two independent carts and setup after transport, collecting images from a live swine eye,
and potentially running a preliminary calibration experiment.
The ROx was fully assembled and operational in the laboratory in the Arizona
Emergency Medicine Research Center (AEMRC).
There are several connections
between the optics cart and the electronic control cart that had to be disconnected and
after transport reconnected: the 2 scanning mirrors, PMT, AOTF, IR driver, and the
Thorlabs power detector all connect from the optical breadboard to the main computer;
the SSM signal connects from the breadboard to the customized electronics, the two
actuators connect from the breadboard to their driver, and the battery pack had to be
connected to the AOTF. The customized electronics are powered by two power supplies,
both of which plug into wall outlets (via extension cords for extended range of motion
within the operating room).
In order to be used in an operating room setting, the optics should be shielded
from their surroundings, particularly miscellaneous fluids and stray light. A cover for the
ROx was designed and implemented using 19 gauge ½-inch mesh welded hardware cloth
(Garden Zone, see Figure 5.1) covered with black ribbed vinyl (WJ Dennis) attached by
screws and washers. The cover slips over the optics and can be secured to the
98
breadboard. It includes a hole in the top for the fan duct to cool the laser, as well as a
flap that provides an escape for the hot air on one side of the laser. A second flap was
added so that an operator can visually spot-check the optical system.
Figure 5.1: Mesh frame for ROx cover, with openings for the eye piece and actuators,
and a protrusion in the back for the IR targeting laser.
An acceptable route was determined from AEMRC to the large animal laboratory
at the far end of the University Medical Center (UMC) across the street. Transporting the
optics cart required two people, lifting the wheels over weather stripping, cracks/bumps
in the sidewalk, and across elevator thresholds. The electronics cart could be managed by
one person. A third cart was used to transport the medical equipment. Upon arrival in
the large animal lab, the electronics were connected and the laser turned on to warm up at
7.96 amps. Trash bags were cut and taped around the electronics cart and the bottom of
the optics cart for protection from fluids, and the data acquisition and illumination control
programs were initiated (Figure 5.5 shows a similar setup). While the other group was
performing their procedures, the powers of all 5 wavelengths were set to ~3μW as
measured by a Thorlabs detector. The optical alignment was then checked by imaging a
99
test target (white paper with printed black lines spaced ~1 mm apart; see Figure 5.2) in
“video mode” with a lens with similar power to the eye. Once the alignment was
confirmed, all wavelengths were then set to ~25μW for imaging the pig eye. No IR was
used in this experiment.
Figure 5.2: Test target; image acquired in trial run.
When the tension pneumothorax experiment was complete, the ROx was moved
into position. However the laser had been running for several hours inside the protective
cover, and though outside air was being blown into the laser, the hot air had nowhere to
go.
The laser overheated and automatically shut off.
While it cooled, a pig was
resituated on its table such that the eye was more easily accessible. Drops were put in the
eye to cause pupil dilation. Additionally, in order to keep the cornea from clouding and
opacifying as the eye is held open, a saline solution (pH balanced at 7.4) had been
prepared beforehand and applied as an eye wash. This solution serves in place of tears
and should be administered at the frequency a human would blink.
100
Figure 5.3: Electronics and optics carts connected and in position to acquire retinal
images.
The laser was turned back on and the powers were checked before acquiring
several images from the eye. However the AOTF had all channels set to 514.5nm, so no
image analysis could be performed. Error! Reference source not found. shows the ROx in
osition for data acquisition. It was noted that there was little or no room for a person to
aim the device in this position. Figure 5.4 shows two of the images obtained. The edge
blanking is clearly visible in both images, as is a reflection from one of the focusing
lenses. Even though the IR was not used, the size and darkness of the vessels indicate
that the left image was almost certainly a vein, and similarly the right was probably an
artery. At 514.5nm, Hb has a higher optical density than HbO2, so veins would appear
darker.
101
Figure 5.4: Retinal vessels imaged at 514.5nm. It appears that the left vessel is a vein
and the right is an artery. The bright region is a ghost reflection from one of the focusing
lenses.
We were unable to track a single vessel for more than a minute or so. This was
due in part to the arrested breathing of the animal (from the prior procedure), particularly
since the operating table, while stable, jerked with every breath. It was decided that the
eye would not be paralyzed, so the eye itself was rolling, as well. In addition, the aiming
method made following even relatively slow eye motion very difficult.
In response to the overheating, a 4 inch secondary fan was added to pull hot air from the
laser out of the covered system. To prevent light from entering the system through the
fan, an 8 foot aluminum dryer duct separates the cover and the fan; the twisting and
curving of the duct effectively prevent stray light from reaching the optical system (see
Figure 5.5). The cooling system tested successfully before the next set of experiments.
C. First Calibration Experiment
a. Setup
A female domestic cross pig was habituated at the animal facilities at University
Medical Center (UMC) for several days and fasted the day before the procedure (no
prohibition of water). In addition to the optics and electronics carts, the medical
102
equipment cart was needed (see Figure 5.5). It contained Foley and Swan-Ganz
catheters, syringes, a centrifuge for Hb, as well as several monitors.
Figure 5.5: Left to right, the medical cart, the optics cart, and the electronics cart in the
large animal lab. The test target is protruding past the eye piece, and the silver hose is to
the second fan which acts to remove hot air from within the black cover. Trash bags are
used to protect the electronics from any fluids involved in the experiment.
The Vigilance monitor displays feedback from the Swan-Ganz catheter (i.e. the
central line, also referred to as the “Swan”), including temperature, SvO2, and cardiac
output (CO). The BioNet Vital Signs Monitor reports heart rate, respiratory rate, and
peripheral oxygen saturation (SpO2). The SpO2 was measured via pulse oximeter placed
on the animal’s snout. The Philips Hemodynamic Monitor reports blood pressures
(systolic, diastolic, and mean arterial pressures), including pressures in the pulmonary
artery. It also reports central venous pressure. The Vigileo Flotrac Montior reports
stroke volume variation (SVV). A fourth cart was also utilized to transport a CO103
Oximeter to measure Hb, HbO2, hematocrit, SO2, and/or SvO2 of blood drawn from the
central line.
As with the trial run, the calibration experiment required transporting the ROx
from AEMRC (Arizona Emergency Medicine Research Center, the location of the ROx
lab) to the UMC large animal lab. Upon arrival, the ROx was reconnected and the Ar++
laser powered on at 7.96 A to warm up. The live pig used in this experiment was a
female domestic cross breed 58 inches in length (nose to tail), weighing 150kg. The
swine was sedated and positioned on its back on the operating table, then anesthetized
using 2% Isoflurane with 1.5 liters of O2 per minute. The operating table used could be
electronically positioned and was significantly more stable than the table in the trial run.
Though the ROx has vertical adjustability, the table had to be adjusted to be within the
range of motion of the ROx. The measurements from the Swan are pressure-sensitive, so
any vertical adjustment of the table required recalibration of the SvO2 monitor. It was
therefore most efficient to prepare the eye and adjust the table height for best aiming
before placing the catheter (“floating the Swan”).
Eye preparation started with administering dilation drops into the both eyes and
taping the eyes shut while the anesthesia takes effect.
Once the pig was fully
anesthetized, a hand-held ophthalmoscope was used to check the clarity and optical
quality of the eyes. The left eyelid was sutchered open. As soon as the tape was
removed and the eyelid opened, a steady regimen of the balanced saline was applied via a
large syringe. The person washing the eye would squirt the eye as frequently as he
blinked; this served as blinking for the pig, and it kept the cornea from clouding. This
eyewash procedure was consistent throughout the duration of the experiment. For a 7.5
104
hour experiment, about a liter of saline solution was needed. Eye preparation lasted
about 30 minutes. The pig was then positioned on the table such that the ROx had ideal
access to the retina. The head was propped up with towels and held in place with a nylon
tether.
Figure 5.6: Pig eye sutchered open for imaging.
While the eye preparation was taking place, the surgical procedure for placing the
catheter was begun (see Figure 5.7). An incision was made in the groin to expose the
femoral artery. After inserting a tube in the artery and tying the artery off, a catheter was
inserted into the artery through the tube.
Placement was confirmed via electronic
waveform output. A similar procedure was performed on a femoral vein. The Swan
Ganz catheter was “floated”: it was inserted through the tube and carried with blood back
to the heart and into the pulmonary artery where the SvO2 can be measured. After the
height of the table was adjusted for aiming the catheter readings were calibrated.
However there were complications with the Swan that remained unresolved throughout
the experiment, so only arterial measurements were made. The entire surgical procedure
(including attempting to resolve Swan complications) lasted 3 hours and 20 minutes.
105
Figure 5.7: Dr. DeLuca performing a femoral cut-down to access the femoral vessels.
After the Ar++ laser had warmed up for an hour, power calibrations were started.
The powers of both the IR laser and the Ar++ laser were set to ~3μW for tests with the
target image. Once the optical alignment was verified and the edge blanks properly
positioned, the powers were equalized at 25μW. The Thorlabs detector was then placed
in the beam path transmitted (“dumped”) by the pellicle in the forward path. This
provided a means of loosely approximating power at the eye, ensuring power never
exceeded safe limits. When the surgical procedure was at a point where the table could
be positioned, the ROx was wheeled to the bedside and aimed using the live streaming
capability of the frame grabber. Minor adjustments were made to the powers such that
the all channels had the same brightness on the computer screen. A few test images were
taken in the blue-green; a razor blade had been inserted into the intermediate image plane
106
between Sph 3 and OAP 4 to potentially serve for image alignment in post processing,
but it was too reflective and was removed before data collection began.
b. Experimental Procedure
The calibration is based on measuring the SaO2 in the femoral artery and
comparing it to data collected with the ROx. The arterial saturation should be the same
throughout the body, so the femoral artery should have the same SaO2 as retinal arteries38,
60
. The SaO2 was controlled manually by adjusting the percentage O2 in the gaseous
mixture the pig was breathing. However there was not a precise way to determine the
FIO2 (Fraction Inspired Oxygen), so instead of using controlled ventilation and allowing
the SaO2 to stabilize for 8 minutes20, the measurements were acquired during “dives”.
When the animal was in a stable condition, the percentage O2 was turned down via the
knob on the gas flow control. The pulse oximeter readout was the primary indicator
used, and when the monitor read out a specific predetermined target SaO2, an image was
snapped with the ROx, the pulse oximeter reading was recorded, and a blood sample was
pulled from the femoral artery and analyzed in the CO-Oximeter (CO-Ox). Each ROx
image was automatically saved with a time stamp, so the images could be correlated with
the other measurements. The time required for a blood draw and CO-Ox analysis was the
limiting factor for the number of ROx-vs.-CO-Ox data points acquired in a single dive.
The pulse oximeter has too much variability to be used for calibration purposes.
This is not the most precise way to calibrate, especially because of the error of the
pulse oximeter and, to a lesser degree, the CO-Oximeter. However the proper gas flow
measurement equipment (which would have allowed for a precise FIO2 stabilization at a
particular SaO2) was not available.
107
For the health of the animal, saturations were dropped for a measurement and
then the percentage of O2 was returned immediately to 100% to minimize possible
damage due to hypoxia. The pig was allowed to stabilize breathing 100% O2 before the
next saturation dive took place. For the health of the animal and the stability of the eye
position, SaO2 levels were not dropped below 69% as read by the pulse oximeter. When
the experiment was completed, the animal was euthanized.
There are multiple varying systematic factors for which compensations had to be
made throughout the experiment.
The image edges drift slightly, the powers and
wavelengths vary significantly, and the pig eye rolls, especially when the SaO2 drops
below 70%. The powers had to be adjusted about every 30 minutes in order to keep all 5
images consistent with one another. The AOTF frequencies were checked when the
powers are adjusted, especially when it is evident that one of the wavelengths had
particularly low intensity despite a higher power setting. The edges might have to be
adjusted every couple of hours. Even with an improved aiming system, it is not practical
to track a vessel when the eye rolls at low saturations.
D. Calibration Data Analysis
Once the data from the calibration experiment was collected, the analysis methods
had to be explored. Since the data set was gathered using the frame grabber, each
interlaced image had to be separated into its respective wavelengths, and each
wavelength identified. Once each image was broken into its 6 sub-images, the vessel and
fundus had to be identified.
This initial analysis was performed by hand on the images converted into Excel
spreadsheets. The user simply approximated the vessel intensity value and the fundus
108
intensity value for each wavelength and then used Equation 41 (variation on Equation 23)
to determine the OD’s. The OD spectrum could then be plotted and used to fit a
parabola. The minimum wavelength at which the vertex of the parabola occurred was
then plotted with respect to the measured time-corresponding SaO2 value as measured by
the CO-Oximeter (Figure 5.8).
 I
OD   log10  blood
I
 reference



   log10  I vessel 

I


 fundus 
41
The plot, though crude, indicates that the minimum of the OD spectrum is
trending with in vivo off-axis intravitreal illumination data.
In looking at the OD
spectrum, it appeared that the parabola could be improved significantly by the omission
of a single outlying point under the premise that the single wavelength was faulty, either
due to systematic noise due to the AOTF or because the laser power at that wavelength
had fluctuated. To this end, each parabola was re-fit to only four of the wavelengths
chosen with the analyst’s best judgment (Figure 5.9).
The slope is closer to the
intravitreal illumination line, but this is clearly not the analysis improvement required for
calibration.
A more thorough analysis is necessary, and some degree of automation is required
for efficiency. A method was developed using averaged raw data to determine the OD of
the blood. For each image, the user would select a vessel and straighten it via drawing a
line down the center of the vessel. The vessel would then be correlated line-by-line as a
secondary potential straightening method and the user selects which method is more
effective.
Once the vessels are straightened, a plot containing all 5 profiles is then
109
displayed and the user must select the region of the profile that is the vessel. The
excluded region is assumed to be the fundus.
Figure 5.8: By-hand calibration analysis compared to the off-axis intravitreal
illumination calibration line in vivo for same SaO2 values (equations shown). The line of
best fit is shown for the first attempt at analyzing each image.
Figure 5.9: By-hand calibration analysis compared to the off-axis intravitreal
illumination calibration line in vivo for same SaO2 values (equations shown) with one
wavelength omitted. The line of best fit is shown for the first attempt at analyzing each
image. The 496.5 nm point was omitted from all fits except for the point corresponding
to SaO2 = 42%, where 476.5 nm was the clear outlier.
110
The values in the vessel region for each wavelength are sorted in order from least
to greatest, and some variable number of the lowest values are averaged and used as the
vessel value. Effectively, only the bottom pixels of the vessel were included. The fundus
value is also determined by sorting, but the determining factor is the gradient between
pixels in the fundus (specifically in the direction of the fast scan). The idea behind this
step is that the fundus profile would be approximately flat (horizontal) in ideal
conditions.
Some variable number of fundus values corresponding to the lowest
gradients is averaged to find the fundus value for each wavelength.
Figure 5.10: Example of set of averaged vessel profiles displayed in order to allow the
user to select the region of the profile that is actually the vessel. The gray box represents
a typical user selection. “Std” indicates standard deviation of the vessel values used;
“Uncertainty” in this case is the standard error of the mean divided by the average vessel
value. The “Highest Std” and “Highest Uncertainty” are pulled from the wavelength with
the greatest standard deviation.
111
The horizontal region was optimized by varying the number of pixels kept from
each row for fundus and vessel values. Similarly, the vertical region of interest (i.e. how
much of the vessel was included in the averaged profile) was also allowed to vary. Of the
region initially selected by the user, both the number of rows used and which row was
used “first” were varied, effectively producing a vertically moving window of variable
size.
To summarize the algorithm, the minimum vessel values were averaged, and the
most spectrally neutral fundus values were averaged. This was performed multiple times
over varying regions. The region of interest was optimized to give the OD spectrum with
the best parabolic fit, using the R2 value of the fit as a metric. Optimizing the region of
interest for vessel and fundus values improved the quality of the calibration line
significantly. The uncertainty in the measurement is quantified by the standard error of
the mean (SEM, where σ is the standard deviation and n is the number of pixels used).
SEM vessel 
 vessel
nvessel
42
Finding the OD via Equation 41 dictates that the uncertainty of each OD value be
calculated as follows: first the ratio of blood to fundus, rbf, should be considered. The
mean pixel values for the vessel and fundus region are represented by xvessel and x fundus ,
respectively.
rbf 
xvessel
x fundus
112
43
Consider that the fundus has a larger mean and a smaller standard deviation than
the vessel, such that the second term under the radical is smaller than the first by at least a
factor of 10. The SEMrbf can therefore be approximated thus:
SEM rbf  rbf
 SEM vessel

 xvessel
2
  SEM fundus 
  rbf SEM vessel
  
 x

xvessel
fundus
 

2
44
For simplicity, the value of SEMrbf/rbf is represented as S. SEMOD is measured in
terms of %OD.
SEM OD 
 SEM rbf

 log10 (S )  1 S   0.4343
S
ln(10)
 rbf

SEM vessel
  0.4343

x vessel

45
This form of analysis resulted in a calibration line that is significantly improved;
however, it has a marked difference in slope from that of the off-axis intravitreal
illumination data.
One user analyzed the entire data set, and the results are shown in the top plot in
Figure 5.11. The same user then repeated the analysis process three more times for each
image. The metric used to determine most reliable data is the R2 value of the parabolic fit
to the OD spectrum, so only R2 values greater than 0.99 were included in a more selective
look at the data. This brought the slope closer to that of the intravitreal line (bottom plot
in Figure 5.11).
While this calibration line clearly trends with SaO2, the OD calculation method is
flawed. When there is a glint in the image, the minimum vessel values are not actually
representative of the center of the vessel. The amount of glint is variable image to image,
adding an element of inconsistency.
113
Figure 5.11: First attempts at automated analysis of the vessel OD. The top graph shows
one user’s first attempt at analyzing the full set of images with the automated process
(optimized ROI). The bottom graph shows the results with an OD spectrum whose
parabolic fit had an R2 value better than 0.99; this includes results from 4 separate
analysis attempts for each image by the same user. Uncertainties are included as error
bars, ranging from 0.6-3.6% SaO2.
Another analysis method is under development, in which a curve is fit to the
vessel profile. The equation for the curve takes into account the vessel radius, laser spot
width and angle of incidence, illumination loss depth, backscatter, and vessel position. A
114
modified nonlinear regression fit is performed to produce a glint-less profile of the vessel
in each wavelength from the available data. The minimum of this curve is used as the
vessel value. The vessel profile is also normalized and adjusted for linear tilt due to offaxis illumination, so the fundus value is simply 1 in this analysis, though the normalized
spectrally neutral region could be used instead.
Since it was effective in the previous analysis method, several iterative processes
were added to optimize the region of the vessel profile used for the fit in attempt to
minimized error due to the glint or other artifacts in the image. Part of the iterative
process includes finding the vessel radius (RV) and the backscatter term and freezing them
such that they are constant throughout all 5 wavelengths while the other terms are
allowed to vary. The vessel radius is not wavelength dependent. The backscatter term is
significantly less wavelength sensitive than the loss depth20, and both have a very similar
affect on the profile shape using the new vessel equation described in Chapter 8.
Initially, the metric used in this iterative process was the R2 value of the parabolic
fit to the resulting OD spectrum. However, upon getting several unbelievable vessel fits,
it was apparent that this was not the best metric for this data set. Instead, the R2 value of
the vessel fit was used. This resulted in better vessel fits, but the OD spectra were now
less believable. Because both metrics should hypothetically produce the same result, the
results of the two metrics are averaged. The standard error of the mean between the two
resulting wavelengths (SEMλ) is used to quantify the uncertainty. The details of this
method are discussed in Chapter 8. The results of the individual methods are plotted in
Figure 5.12, and the results of the average are shown in Figure 5.13.
115
Figure 5.12: Comparison of calibration results using two different metrics for vessel
curve fitting: the R2 value of the parabolic fit to the OD spectrum (left) and the R2 value
of the vessel curve fit (right). Both plots use only the first attempt at vessel analysis.
Note that these are only from the first pass analyzing the vessel with these
methods, and that both analyses were performed on the exact same image segments and
vessel profiles. Both of these methods show a marked improvement over the top graph in
Figure 5.11.
The slopes are significantly closer to that of the off-axis intravitreal
illumination line, and the R2 value of the linear fit is better, as well. I expect that the
average of repeated analyses would yield further improvement in the results, as seen in
the first automated analysis (bottom of Figure 5.11).
The Figure 5.13 error bars correspond to the SEMλ of the averaged wavelength
values. Using the off-axis intravitreal illumination calibration line as the example, the
uncertainty in %SO2 corresponds to a SEMλ by a factor of ~320, 47. An uncertainty of less
than 3% SO2 requires the uncertainty of the wavelength corresponding to the minimum in
the OD spectrum to be less than 1 nm. The bottom graph in Figure 5.13 retains only the
points that meet this standard. Note the further improvement in the slope towards the
expected off-axis intravitreal illumination calibration line. The R2 value also improves
significantly, but there are fewer points in the plot.
116
Figure 5.13: Calibration lines from combining results from the two analysis metrics. The
top graph shows all data points, where error bars are the SEMλ. The bottom plot uses
only points for which the SEMλ is less than 1 nm, corresponding to an uncertainty of
about 3% SaO2.
It is important to note the shift of the calculated ROx data to the right from the
expected off-axis intravitreal illumination line. The wavelength-dependent effects of
scattering in whole blood in the blue-green spectrum have been documented in prior
117
work17, 19, 20. In transmission, scattering linearly shifts the OD spectrum of whole blood as
compared to pure Hb; it is an extra source of loss of light and creates a spectral shift
towards red19. Compared to the OD spectrum of pure Hb in transmission, the OD
spectrum of whole blood in transmission shifts linearly towards higher wavelengths,
whereas data from off-axis intravitreal illumination demonstrated that the OD spectrum
shifts linearly toward lower wavelengths in reflection. It is reasonable that the inverse
relationship between reflection and transmission would result in a change in the sign of
the slope; scattering from whole blood in reflection increases as wavelength increases20.
However, the data from this calibration experiment behaves to the contrary. The
ODs are higher (i.e. there is more loss from the vessels), and the OD spectrum shifts
towards higher wavelengths. In fact, the shift is greater than that produced by whole
blood in transmission. One possible explanation of this difference is the field stop wire
filter (FSWF). This pinhole with a bisecting wire should be placed at the conjugate
image plane immediately before the PMT, and it should be centered such that the
specular reflection (i.e. unscattered light) is blocked from the PMT. If the wire is not
centered properly, the desired scattered light is blocked, and specularly reflected light is
allowed to pass. Blocking the scattered light would create a loss instead of a gain (since
the specular reflection is ignored for OD calculations anyway) and possibly invert the
sign of the slope of the scattering component.
Another explanation might be that the FSWF could have been aligned such that
the backscattered light could have been blocked, and the light that is detected in the
return path is light that has passed through the vessel and is reflected off of the fundus—
i.e. single-pass light in transmission instead of backscattered light (in reflection). This
118
would explain the directionality of the shift and is supported by recent work by Rodmell
et al.
Due to oversight, the alignment of the FSWF was not double-checked before this
experiment, and it could have been jostled in one of the moves to or from the large
animal lab at UMC. This large shift is not clearly apparent in the data from the first 5 pig
experiments, and it does not seem to be present in later experiments, either. However,
when testing different FSWF sizes, a large shift in the opposite direction was observed in
one set of human data (Chapter 7); this indicates that the size and position of the FSWF
probably affects the scattering term for this system.
In short, this is an important issue and should be kept in mind as the project
continues. Fortunately once the system is set in permanent state, the scattering behavior
should be consistent and a useful calibration line can be measured.
E. Vessel Diameter Measurements
In this process of fitting the vessel profiles, the diameters of the vessel segments
were determined by taking the average vessel radii from the first iterations of the fits.
The units in this case are pixels, and because the scan mirror amplitudes are not precisely
determined, an exact measure in units of μm is not possible with this data set. However,
it is reasonable to approximate that the scan region on the retina was something like 2.5 x
2.5 mm. The images from Pig 6 (this calibration experiment) were collected via the
frame grabber board, so the width of each pixel in this case corresponds to ~3.8 μm. My
best estimate of vessel widths calculated by the fit process is 45-85 pixels or ~170-320
μm.
119
Figure 5.14: Example of vessel profiles corresponding to the vessel selection in the
bottom right image and its calculated width. The image is the sum of all 5 wavelength
images co-aligned. The dark region at the top of the image is caused by vignetting. The
region selected is an artery, but a vein can be seen in the bottom left corner of the image.
The diameter determined by the vessel fit metric is on average 12 pixels wider
than the diameter as determined by the parabolic spectral fit metric. Due to the differing
natures of the analysis metrics, it seems reasonable to trust the diameter values from the
vessel fit metric analysis as opposed to the average between the two analyses or those
from only the analysis using the parabolic spectral fit as the optimization metric. For
different regions of a given vessel imaged throughout the latter half of the experiment,
the vessel diameter is 66 pixels wide, on average, with SEMRv = 3 pixels. This comes to
120
an approximate vessel diameter of 250 μm, which is in the middle of the expected range
of vessel widths.
One example image and its corresponding vessel profiles and
determined width is shown in Figure 5.14. Note that the vessel fits are resultant of the
vessel fit metric analysis.
Measuring the vessel diameter not only improves the fit to the vessel profile, but
it is useful in determining the path length of light that has interacted with the vessel49.
This is directly related to the scattering term in whole blood which necessitates
calibration of the ROx in the first place. Vessel diameter calculations could be also
useful in further studies of the vasculature and metabolism in the eye. The vessel
diameter is related to the amount of light scattered by the blood, where greater scattering
occurs with larger vessels48. An inverse relationship between vessel diameter and SO2
has been shown that in cases of hyperoxia61, 62 as well as possibly vessels with diameters
< 90 μm63.
F. Conclusions from First Calibration Experiment
The first calibration experiment, while very useful, left much to be desired
experimentally. It is evident that more control is required when changing the amount of
O2 inspired by the animal. The animal’s physiology should be allowed to stabilize for a
given percentage inspired O2, and better image quality (specifically, higher resolution and
bit depth) is required. The poor aiming capabilities of the ROx were also a hindrance.
All of these issues were addressed and improved upon for the next set of pig experiments:
a more precise gas mixer is available to the physicians and anesthesiologist, allowing for
more control of the inspired O2. The frame grabber has been replaced with the PicoScope
121
to improve image quality, and the ROx is now mounted on a jack with 5 degrees of
motion.
Ideally, the data would have produced a calibration line with a slope of 3.03 in
keeping with previous work, and the majority of data would be useful in the first pass of
automated analysis. Though the data from this experiment did not produce the quality of
calibration line expected, it did allow for significant improvement to the analysis
methods.
G. Second Calibration Experiment
Figure 5.15: ROx in the pig lab after several improvements, including the jackmount for
the optics breadboard. Optics are uncovered here for pre-experiment alignment and tests;
the same black cover was put in place before imaging the pig eye.
122
Several improvements were made to the ROx, including the jack mount for better
aiming and the implementation of the PicoScope for improved image quality (increasing
both spatial and pixel depth resolution). With a protocol revision, the experimental
procedure was made more robust with plans for performing both calibration and sepsis
experiments on the same swine in the same experiment and contingencies to allow for
only one of the experiments to be performed if something is preventing the completion of
another. A means of paralyzing the eye was included to avoid the issue of the eye
rolling, especially at low SO2. Other improvements included a more precise means of
controlling the %O2 inspired by the swine. With these changes completed, a second
calibration experiment was attempted.
a. Setup and Procedure
As in the previous experiment, a female pig was habituated for several days and
fasted the day before the procedure (no prohibition of water). The ROx (optics on auto
jack and electronics on rolling cart), the hemodynamics cart, and the CO-Ox cart were
wheeled to the large animal lab at UMC. The pig was anesthetized and intubated the
morning of the procedure as before. However, this animal was significantly smaller than
the previous six, weighing only 31 kg with a nose-to-tail length of 35 inches. Ideally, the
lines would have been placed, the calibration experiment run, and then the sepsis insult
would begin. In this case, though, the size of the animal made the placement of the Swan
and the arterial lines particularly difficult. After struggling with the lines for an hour with
no success, sepsis was induced via cecal ligation and perforation (described in Chapter 6)
in order to ensure that at least some useful information was gathered from this
123
experiment.
Eventually, both lines were successfully placed, allowing for accurate
measurements of SaO2 and SvO2.
During the surgical procedure on the abdomen, the left eye of the animal
underwent preparation.
Because of the issue of the eye rolling in the previous
experiment, paralysis of the eye was deemed necessary. The paralysis procedure chosen
is the sub-Tenon’s administration of a local anesthetic64, 65. It is cost effective, straightforward, and does not require much extra equipment. Lidocaine was the anesthetic of
choice, and 2 mL were injected into the sub-Tenon of the pig’s left eye (Figure 5.16).
The sub-Tenon is located between the sclera and the Tenon’s capsule (or the bulbar
sheath), which forms the eye socket. In addition to the paralysis, the eyelashes were
clipped, dilation eye drops were administered (1% cyclopentolate hydrochloride), the eye
was sutchered open, and a frequent wash of the 7.4 pH-balanced saline solution was
consistently applied.
Figure 5.16: Paralysis of the eye via injection of lidocaine into the sub-Tenon region of
the pig eye. Eye clamps are used to hold the eye open.
124
Retinal imaging with the ROx began as soon as the arterial line was placed. The
image quality was significantly improved with the implementation of the PicoScope and
a reference power detector (Figure 5.17). Whereas a capacitor was used to effectively
blur and smooth lower resolution images with the frame grabber in previous experiments,
the PicoScope produces crisper images and removes artifacts due to power fluctuations
(such as ringing, seen as vertical lines on the right edge of the frame grabber image in
Figure 5.17). Unfortunately that also included resolving the central glint on the vessels.
Note that both images in Figure 5.17 are all 5 wavelengths co-aligned and stacked; the
edge blanking is not well aligned between wavelengths, resulting in the gray edges in
both images.
Figure 5.17: Comparison between images collected by the frame grabber board (left) and
via the PicoScope (right). These images are the sums of all 5 wavelengths co-aligned.
There are reflections off of the focusing lenses in both images, but the resolution
difference is clear. Also note that ringing (vertical lines on the right side of the frame
grabber image) is not present in the PicoScope image. The image on the left is from the
first calibration experiment, and the image on the right is from the second.
In an attempt to block more of the glint, the field stop wire filter (FSWF) was
switched mid-experiment, doubling the size of the bisecting wire for the same 2 mm
pinhole. However, the alignment could not be thoroughly tested in a time-efficient
125
manner, so the glint problem was not improved. The eye quality also degraded somewhat
during the time it took to change the FSWF. Note that the glint was present, and that the
degree of variation of the glint size in a data set was independent of pinhole. Glint is less
apparent when the vessel is out of focus.
Blood was periodically drawn from both the femoral artery and vein to be
analyzed by the CO-Oximeter.
The ROx was used to collect retinal images
corresponding to the draw times. The IR laser was not available due to alignment issues,
so the size, placement, and darkness of the vessels in the blue-green were used to attempt
to distinguish between arteries and veins. In this case, an artery-vein pair should be
found (common near the optic nerve head), and the desired vessel can be followed away
from the optic nerve head.
Figure 5.18: Comparison of glint from vessel images acquired with a 2 mm pinhole
bisected by a 200 μm wire (left) and by a 400 μm wire (right).
126
Calibration data were gathered after the onset of sepsis (the second FSWF was in
place), beginning about an hour before the anticipated end of the experiment. The
inspired O2 was lowered from 32% to 18%, and then lowered again to 17.5% before
raising it as the pig was euthanized. The animal’s physiology was allowed to stabilize for
several minutes before a blood draw and retinal image were made. The SaO2 was not
constant, however, probably due to the severe septic state of the pig by that point in the
experiment.
b. Data and Analysis
Most of the arterial blood draws run through the CO-Ox were taken with the
second FSWF in place, so that data is presented as the rough “calibration” data. The
placement of the FSWF was not tested rigorously due to time constraints of the
experiment, so there is not a good means of comparison between data taken with the
different filters. Therefore, for calibration purposes, data from the two filters cannot be
compared due to the potential sensitivity of the scattering component to the position of
the FSWF, as discussed in Section F.
The data set from this experiment was analyzed a few ways.
First, it was
analyzed by hand using the GUI described in Chapter 8. Three people analyzed the data;
they had varying degrees of experience with the data (8 weeks to 2 years on the project),
and only the most experienced produced a calibration line with any correlation at all. The
calibration by hand in Figure 5.9 is better. While a very experienced individual could
doubtless produce a plausible calibration line this way, it would be incredibly time
consuming.
127
The automated analysis detailed in Chapter 8 was also performed on the data.
Figure 5.19 shows the results from the average of the wavelengths corresponding to the
minimum OD found using the two iterative processes with differing metrics (vessel curve
fit and spectral parabolic fit). Again, this combination produces a calibration line with a
better R2 value and slope closer to that of the intravitreal illumination calibration line than
either iterative process individually. Also note the position of this calibration line with
respect to the intravitreal calibration line; it is significantly more similar than in the
previous experiment. This could indicate that the scattering in this setup is similar to the
scattering in the off-axis intravitreal illumination experiment (e.g. more reflected spectral
information) and/or that the shift in the previous calibration experiment was anomalous.
Figure 5.19: Calibration line from combining results from the average analysis metrics,
where error bars are the SEMλ. Note the scale is different than that of Figure 5.13. The
error bars range from 0.99-4.27 nm, corresponding to an uncertainty of about 3-13%
SaO2.
Note that the scale of Figure 5.19 is different from the previous graphs. The error
bars range from 0.99 to 4.27 nm (SEMλ), corresponding the SO2 uncertainties of 3-13%.
128
The average SEMλ for first-pass data from the first FSWF (2.4 nm) is slightly lower than
that of first-pass data from the second FSWF (3.1 nm). Both are significantly greater
than the average SEMλ from the previous calibration experiment (1.6 nm).
c. Conclusions from Second Calibration Experiment
It was predicted that the higher resolution afforded by the PicoScope would
improve the image quality and therefore improve the quality of the calibration
measurements. Instead, the increased clarity of the glint in the new images has created
challenges for data analysis and consequently calibration ability.
This experiment did confirm that image quality with the PicoScope is much
improved in terms of resolution and image noise. The paralysis of the eye was a
successful and useful addition to the experiment protocol, and the improvement to the
aiming capabilities due to the new jack mount was evident to the ROx operators.
However, this experiment also made the shortcomings of the FSWF as a glint
blocker evident. The 200 μm wire was too small, corresponding to a spot on the retina of
about 30 μm (smaller than the size of the incident spot on the retina in some
wavelengths). The precision of the position of the FSWF after switching to a 400 μm
wire was not satisfactory, so either the increase in wire diameter was not enough, or the
placement of the filter was incorrect (or both). In any case, the glint is larger than the
automated analysis program is designed to handle. The automated analysis has been
shown to be effective for vessel profiles with little or no glint, and further analysis and
testing of the effectiveness of the FSWF should be pursued. A set of experiments to this
end are described in Chapter 10.
129
H. Conclusions
It is necessary for the ROx to be calibrated in order for it to be used as a
diagnostic instrument during the onset of sepsis (or any other shock condition) in live pig
experiments or for any human testing. The ROx has been used to obtain retinal images
from seven live swine, two of which were used, at least in part, for calibration
measurements. The ROx was successfully transported to and from the animal lab, retinal
images were acquired, and corresponding %SaO2 data was recorded from blood draw
from femoral vessels and measured using a CO-Oximeter.
The retinal images were analyzed via several different methods, the most
consistent and accurate of which was an automated analysis that fit theoretical vessel
profiles to actual vessel data in all 5 wavelengths to determine the OD spectrum of the
retinal blood. The dual-metric analysis process was used to determine the wavelength
corresponding to the minimum OD value in the blue-green spectrum of the blood.
Despite the improved analysis process, the ROx remains uncalibrated.
The
resolution and pixel depth of the images from the first calibration experiment limited the
quality and allowed for unacceptable error. There is also a linear shift in the calibration
line that is contrary to the expected results (using the calibration line from off-axis
intravitreal illumination) and not well understood.
Though there are possible
explanations (including the position of the FSWF), it was not seen in the second
calibration attempt. One possible explanation could be that, while the off-axis intravitreal
illumination data represents the OD spectrum in reflection20, the placement of the FSWF
may block more of the backscattered light and allowed for detection of single-pass light
that has passed through the vessel and is reflected/scattered by the fundus back into the
130
collection path. The shift observed in Figure 5.13 is larger than the shift seen in the OD
spectrum of whole blood in transmission (as compared to Hb)19. This is also unexpected,
indicating a potential problem within the collection path or an aspect of the system that
has yet to be explored or explained (e.g. the eye acting like an integrating sphere,
increasing the presence of the scattering spectrum into the OD spectrum, explained
further in Chapter 10).
The second calibration attempt, while demonstrative of several improvements to
the ROx, was not successful due to the dominant presence of the central glint in the wellfocused vessel images. It was difficult to determine the true vessel intensity value,
making it difficult to accurately calculate the OD spectrum and utilize the BGO technique
for determining %SO2. Though the calibration line is closer to that found using off-axis
intravitreal illumination, the uncertainty of the data is significantly worse.
Though insufficient for a true calibration, the robust nature of the BGO has
become ever more apparent as both of these data sets produced calibration lines that trend
closely with the calibration equation from the off-axis intravitreal illumination data. The
amount of uncertainty is undesirable, but with improvements to glint removal, there is
potential for increased confidence in the SO2 measurements. The glint can most probably
be removed or significantly reduced by further development of the FSWF. Once this
upgrade is completed, there is a high probability that a successful calibration experiment
can be performed, resulting in an accurate and consistent calibration can that can be used
with the ROx in future experiments.
131
Chapter 6:
SEPSIS EXPERIMENT ON LIVE SWINE
A. Introduction
One of the goal uses for the ROx is earlier (and noninvasive) detection of sepsis.
The device was used in a preliminary set of experiments in which sepsis was induced in
live anesthetized swine. This chapter describes these experiments. Section B outlines the
motivation and background behind the sepsis experiments. Section C describes the
experimental setup and procedure sepsis experiments. Section D contains data analysis,
and Sections E and F give results and conclusions.
B. Sepsis and the ROx
Sepsis is a life threatening medical condition in which the blood is infected, which
can lead to organ dysfunction and death. It is defined as a condition in which the patient
has SIRS (Systemic Inflammatory Response Syndrome—systemic manifestations of
infection) and documented or suspected infection. When organ dysfunction or tissue
hypoperfusion occurs in addition to sepsis, the condition becomes severe sepsis. In its
advanced stage, septic shock, a patient with severe sepsis no longer responds to fluids66.
Infection is introduced to the blood stream, often through the abdomen, lungs, or
urinary tract. The bacteria/toxin causes vasodilation of the blood vessels, which results in
lower venous pressure and consequently hypovolemia and lower systemic vascular
resistance (SVR)67. In other words, the vessels enlarge without flow increasing, so
relatively less blood is available and more oxygen is pulled from the arteries due to the
132
increased surface area. This can be seen as a drop in venous SO2 and venous blood
pressure.
Sepsis is a condition in which speed of treatment is critical. In 2001, Rivers
showed that Early Goal Directed Therapy (EGDT—a protocol-based treatment begun in
the emergency room as opposed to standard therapy using the “art” of medicine in the
ICU) improved the in-hospital mortality rate from 46.5% to 30.5%68.
One of the
differences between EGDT and the standard therapy is that EGDT uses the central
venous pressure (CVP) and central venous SO2 (ScvO2) via use of a central venous
catheter (CVC) instead of measuring pulmonary capillary wedge pressure and the mixed
venous SO2 (SvO2) via a pulmonary artery catheter (PAC). Though there is less risk in
placing a CVC than a PAC, a noninvasive method of determining SO2 and venous
pressure would further improve EGDT by decreasing time and potential hazard involved.
In fact, there are cases in which sepsis is introduced by placing the central line.
Previous work shows that peripheral venous pressure (which can be measured via
peripheral IVs) is a viable and less invasive alternative to CVP69. It has also been shown
that retinal venous SO2 (SrvO2) tracked closely with the SvO2 during the development of
hypovolemic shock and resuscitation thereof60, 70, 71. It is therefore reasonable to further
test the correlation between SvO2 and SrvO2, particularly during the onset, progression,
and resuscitation of sepsis. Because of the potentially rapid and noninvasive nature of
data collection of the ROx, BGO retinal oximetry has the potential to be used to measure
SrvO2 and replace the CVC (when used in conjunction with peripheral IVs). If the CVC
can be replaced with less invasive or noninvasive means, the widespread use of EGDT
could be more strongly justified. It is therefore important to attempt to measure SrvO2
133
with the ROx during the onset, progress, and resuscitation and compare the
measurements to those of SvO2.
C. Sepsis Experiment
As mentioned in Chapter 5 on the calibration experiment, animal testing is
necessary at this point in the research process. Pigs are the best analog to humans, as
they are the lowest species with acceptably comparable vessel size, cardiac anatomy, and
eye size and quality. They also have similar physiologic responses to sepsis and the
resuscitation thereof in humans.
While rats serve as an acceptable analog for the
physiological aspects of the experiment, it would not be practical to attempt to gather
data with the ROx during a sepsis experiment on a rat. In order to comply with IACUC
standards, the experiment is designed to use the fewest number of pigs in the most
painless way possible for the animals. Another consideration was the time and cost of the
experiment.
In order to use the large animal lab at University Medical Center, a
veterinarian and a veterinarian technician must be present, and the cost of their labor
must be considered. There is also a cost for housing the animals once they arrive.
Previous methods of sepsis induction (including cecal ligation and puncture) lack
consistency and take at least a day for severe sepsis and septic shock to set in 72-75. These
lengths of time are not acceptable for the cost limitations, so a new sepsis model is
required.
A sepsis model that is useful for these experiments should cause the onset of
sepsis, severe sepsis, and septic shock within 8 hours, including any preparation or
operating time. The animal must be able to be resuscitated from septic shock, however.
134
The different stages need to be observable, as well. The model, in short, must be
controllable and repeatable.
One sepsis model that has been explored with minimal success involved
implanting feces into the peritoneum directly (where the abdominal organs are located)76,
77
. However, this method alone does not sufficiently induce sepsis. Cecal ligation is a
procedure in which the abdomen is opened and part of the cecum (the beginning of the
large intestines) is isolated and tied off to deprive tissue of oxygen, creating an ischemic
insult. It has been used to induce sepsis with some success in swine 78. Cecal ligation can
be used in combination with a bowel perforation model in which cecum is punctured or
perforated in order to let bacteria and fecal matter within the cecum leak into the
peritoneum. This combination creates the infection that is currently a standard model of
sepsis75.
The first method tried by this group was cecal ligation and puncture, but a fecal
slurry was added as a preliminary boost to the infection. To minimize distress, pain, or
suffering, the animals were anesthetized during the entire experiment, including
euthanization. The swine was sedated via injection prior to the start of the procedure, and
then intubated and anesthetized for the duration of the experiment.
To maintain
consistency of oxygen inhaled, the pig was put on a ventilator which maintained a steady
gas volume and rate of breathing.
a. Setup
For the first 4 pigs (after the initial trial run described in Chapter 5), the
experimental setup is nearly identical to that described in Chapter 5.
The optics,
electronics, and medical equipment carts, as well as the CO-Oximeter/printer cart, were
135
transported from AEMRC to the UMC large animal lab. Upon arrival, the ROx was
reconnected and the Ar++ laser powered on at 7.96 A to warm up. The live pigs used in
these experiments were female domestic cross breeds between 124 and 150 cm in length
(nose to tail), weighing between 50 and 60 kg.
The sixth sepsis experiment was
performed on the much smaller pig (described in Chapter 5), which introduced
complications to the procedure. The ROx had also been updated with the jack mount and
the PicoScope for the last experiment.
For the purpose of inducing sepsis, the animal lab staff collected fresh feces from
the pig before it was sedated, for use in the fecal slurry. The swine was sedated and
positioned on its back on the operating table, then anesthetized using 2% Isoflurane with
1.5 liters of O2 per minute. The operating table used could be electronically positioned;
though the ROx has vertical adjustability, the table had to be adjusted to be within the
range of motion of the ROx. As mentioned before, the measurements from the Swan are
pressure-sensitive, so any vertical adjustment of the table required recalibration. It was
therefore most efficient to prepare the eye and adjust the table height for best aiming
before placing the catheter.
Eye preparation started with administering dilation drops into the both eyes. It is
imperative for the eyes to remain moist throughout the entire procedure, because the
cornea will opacify and degrade the optical quality if it dries out. In the first couple of
experiments, the eye was open slightly after the dilation drops were applied, causing a
translucent line across the eye. To prevent this, the eyes were taped shut with bandage
tape when the pig was first sedated, and then re-taped after the drops were applied so that
the eye remained shut while the anesthesia took effect and the surgery was completed.
136
Additionally, paralysis of the eye was implemented for the sixth experiment using the
sub-Tenon’s method described in Chapter 5.
Once the pig was fully anesthetized, a hand-held ophthalmoscope was used to
check the clarity and optical quality of the eyes (see Figure 6.1). The left eyelid was
sutchered open. As soon as the tape was removed and the eyelid opened, a steady
regimen of the balanced saline was applied. For the first couple of pigs, a drip system
was set up such that a drop of pH balanced saline went into the eye approximately every
two seconds (as in Figure 6.2). However the drop did not sufficiently wash the whole
eye, so the cornea clouded up where the drops were not reaching. The answer was
having a person wash the eye via large syringe. The person washing the eye would squirt
the eye as frequently as he blinked; this served as blinking for the pig, and it kept the
cornea from clouding. This eyewash procedure was consistent throughout the duration of
the experiment.
Figure 6.1: Checking optical quality of eye as well as becoming familiar with location of
retinal vessels for aiming purposes.
137
Eye preparation lasted about 30 minutes. The pig was then positioned on the
table such that the ROx had ideal access to the retina. The head was propped up with
towels and held in place with a nylon tether (Figure 6.2).
Figure 6.2: Anesthetized pig positioned for imaging: eye is sutchered open with saline
drip line (not used in later experiments). Note: the drip line is not connected to saline bag
yet.
b. Surgical Procedure
While the eye preparation was taking place, the surgical procedure was begun. In
order to make the best use of time, the procedure for inducing sepsis was performed first.
The abdomen was opened down the midline, and the bladder was identified.
All
incisions were made with an electrocautery pen. The urinary output is indicative of the
stage of sepsis, so a Foley catheter connected to a graduated collection container was
inserted into the bladder. The cecum was then identified (Figure 6.3 left), and a few of
the cecal vessels were isolated and tied off (ligated, see Figure 6.3 right), inducing
138
ischemia in the cecal tissue, and then a 1 cm incision was made in the cecum to allow
leakage of fecal matter. A liquid fecal slurry was then prepared, consisting of the fresh
fecal matter mixed in water. 1 gram of feces was used per kg the pig weighed. The
slurry was then poured into the open peritoneum. The abdomen was then closed in order
to allow the sepsis to progress.
Figure 6.3: Left, Dr. DeLuca has identified the cecum. Right shows cecal ligation.
As with the calibration experiment, an incision was made in the groin to expose
the femoral vessels. After inserting a tube in the artery and tying the artery off, a catheter
was inserted into the artery through the tube. Placement was confirmed via electronic
waveform output. A similar procedure was performed on a femoral vein. The Swan
Ganz catheter was “floated”: it was inserted through the tube and carried with blood back
to the heart and into the pulmonary artery where the SpO2 can be measured. After the
height of the table was adjusted for aiming the catheter readings were calibrated.
139
Figure 6.4: Pig sewn up after cecal ligation and fecal slurry insult and placement of
Foley catheter.
Throughout the surgery and experiment, the animal’s vital signs were recorded (as
available) every ten minutes.
The Surviving Sepsis Campaign66 recommends MAP
and/or CVP target values of ≥ 65 and 8-12 mm Hg, respectively, as part of the treatment
protocol for severe sepsis. In order to allow for the development of severe sepsis, the
MAP was maintained between 50-60 mm Hg after the initial onset of hypotension (MAP
≤ 60). The septic state takes approximately 2.5 hours to develop. In the mean time,
fluids and pressor agents were administered as needed to maintain a mean arterial
pressure (MAP) of greater than 60 mm Hg. “Fluids” consisted of a bolus of normal
saline, ranging from 250 mL to 1 L at a time. Over the course of these experiments, it
was determined by Dr. DeLuca that only a single 250-500 mL bolus should be
administered during this time to expedite the onset of sepsis. Fluids are administered to
counteract hypotension, and pressors are used to help maintain vascular pressure.
Basically, fluids and pressors are used to control cardiac output.
140
Once the SvO2 reached 50-60%, the animal was considered to be in septic shock
and was resuscitated according to a modified Rivers EGDT protocol. Boluses of normal
saline were given repeatedly until the central vascular pressure (CVP) is up to 12-15 mm
Hg, ideally. 20 mL/kg was determined to be the optimum bolus. A drip of epinephrine
(or other another vasopressor) was used to maintain MAP > 60-65 mm Hg. Additional
boluses of epinephrine or dobutamine were administered if the cardiac output was low or
SvO2 is persistently lower than 65-70%.
Vitals continued to be recorded every 10
minutes. Blood samples were taken periodically in order to measure SvO2, central venous
SO2 (ScvO2), and/or peripheral venous SO2 (SpvO2).
The ROx also took periodic
measurements, ideally simultaneously, representing SrvO2.
In the event of a non-
perfusing arrhythmia (which did occur in one pig), the animal is resuscitated using
standard protocols including electrical cardiac stimulation (e.g. cardioversion,
defibrillation, pacing) and administration of appropriate medications as necessary.
With the exception of one experiment, the animal was always euthanized within 7
hours of perforation of the cecum.
The exception occurred in the June 20, 2012
experiment (Pig 3): the pig was rolled into the prone position in order to image from the
other eye.
However, this motion changed the calibration of several hemodynamic
measurements, and vitals ceased to be recorded at that point. The experiment lasted
about 2 hours longer than the rest.
In all cases, the animal remained anesthetized
throughout the experiment and euthanization.
Another exception of note is the July 26, 2014 experiment (Pig 7). There were
several unforeseen variables with this experiment. This experiment was meant to serve
as both a calibration experiment and a sepsis experiment, which is why it is included in
141
both chapters 5 and 6. However, due to complications in placing lines, sepsis was
induced first and the lines were placed later. The pig was a significantly smaller animal
than expected, which means the vasculature was tiny and should have required smaller
surgical equipment. It took roughly 2 hours longer than normal before all hemodynamic
readings could be made. This meant that the calibration experiment was not performed
as planned (i.e. before septic insult).
As discussed in Chapter 5, the field stop wire filter (FSWF) was changed out midexperiment, which introduced another variable to the experiment.
The veterinary
anesthesiologist working with the group on Pig 7 was an addition to this experiment, and
she drew from prior experience while handling the anesthesia and inspired gases.
However, this adds an unanticipated variable to this specific experiment. Her knowledge
and experience will certainly benefit future work once the goals and protocols are more
clearly communicated.
c. ROx Data Acquisition
After the Ar++ laser had warmed up for an hour, power calibrations were started.
The powers of the laser were set to ~3μW for tests with the target image. Once the
optical alignment was verified and the edge blanks properly positioned (controlled via
potentiometers), the powers were equalized at 25μW. The Thorlabs detector was then
placed in the beam path transmitted by the pellicle in the forward path without entering
the system, providing a means of loosely approximating power at the eye and ensuring
power never exceeded safe limits. When the surgical procedure was to the point at which
the table could be positioned, the ROx was wheeled to the bedside and aimed using the
live streaming capability of the frame grabber (Figure 6.5 and Figure 6.6).
142
Minor
adjustments were made to the powers such that the all channels had the same brightness
in the video on the computer screen. Note that the mount for the optics in Figure 6.5 and
Figure 6.6 was changed twice over the course of these experiments: first to a spring
suspension mount, and later to the jack mount discussed in Chapter 4.
Figure 6.5: Entire experimental setup: swine after septic insult and intubation and ROx in
position to acquire retinal images.
Images were collected from the eye. The protocol for timing of image acquisition
evolved as the group became more experienced. Images from the first few experiments
were collected at random as the ROx operators saw fit. This continued to some degree
with later experiments, but acquisition timing was also coordinated with changes in
inspired O2 and/or blood draws to be analyzed by the CO-Ox. It would be better in future
143
experiments to snap images every 10 minutes (in coordination with vital sign recordings)
in addition to acquisitions coordinated with blood. This would improve correlation
analysis between SrvO2 and other hemodynamic and vital sign data.
Figure 6.6: The operator’s view of the ROx. The computer CPU, oscilloscope, actuator
driver, and slow scan signal generator are visible on the electronics cart. The back of the
optics cover can be seen, showing the open flap by the Ar++ laser (to the left) and the
room allowed for the IR laser position.
D. Results
a. Sepsis Model
Data is combined from pigs 2-5 and 7 as available. Pig 1 was the trial run, and
Pig 6 was the calibration experiment; neither included septic insult, so they are not
considered.
In all five sepsis experiments, the animal developed septic shock and
144
survived until the end of the experiment. The data for each experiment is organized by
time elapsed from the closure of the abdomen following the septic insult.
Figure 6.7 shows the progression of temperature and heart rate throughout the
experiments. Both show an increasing trend over time. This is congruent with the onset
of SIRS; the temperature increases as the body fights the infection, and the animal
develops tachycardia (heart rate increases) as the body tries to deliver more oxygen by
delivering more blood.
Twice pigs required approximately 1 minute of chest
compressions, and Pig 4 additionally needed cardioversion for tachydysrhythmia (i.e. the
heart was arrhythmic and beating too quickly, so the animal received electric
stimulation). These occurred between vital sign recordings, so they do not show up on
the heart rate plots.
Because this septic model was untested, the medical team was more cautious with
the first few animals, allowing sepsis to develop slowly and less severely, and
resuscitating the animal (still under anesthesia) before euthanization. However, with
growing experience and confidence in the model and control thereof, the infection was
allowed to progress further and faster. This can be seen in the progression of temperature
throughout the different experiments.
Figure 6.7: Progression of temperature and heart rate during the progression of sepsis,
where time 0:00 represents closure of the abdomen after septic insult.
145
Figure 6.8: Progression over time of Mean Arterial Pressure (top left), Central Venous
Pressure (top right), Cardiac Output (middle left), Systemic Vascular Resistance (middle
right), Stroke Volume (bottom left) and mixed venous Oxygen Saturation (bottom right).
Time 0:00 represents closure of the abdomen after septic insult.
Another set of measurements made throughout each experiment includes vascular
pressures (mean arterial pressure, or MAP, and central venous pressure, or CVP), cardiac
output, and the resulting systemic vascular resistance (SVR). SVR is effectively the
resistance that the heart pumps against, and it is calculated using Equation 46, where 80
is a unit conversion factor.
146
SVR 
80MAP  CVP 
CO
46
These measurements are plotted in Figure 6.8. These values are used in the
classification of shock; specifically, they can be used to identify septic shock79. At the
onset of sepsis, bacteria toxin causes vasodilation, resulting in an increase in SVR and a
drop in CVP. In septic shock, the preload (stroke volume, SV) and afterload (SVR) on
the heart decrease, but the cardiac output increases. The systemic oxygen delivery
increases, but the tissue is not properly extracting oxygen (i.e. extraction rate is down), so
the SvO2 also increases. After a time, the body can no longer keep up due to lack of
oxygen, so the cardiac output and SvO2 drop.
Time (in minutes)
Mean
Range
MAP < 60 mm Hg
122.8
40-250
SBP < 90 mm Hg
76
40-120
SvO2 < 70%
104.4
35-220
SvO2 < 60%
134.8
45-260
Table 6.1: Mean times and ranges (in minutes) from perforation of the cecum to
endpoints in the sepsis model (left column). SBP is Systolic Blood Pressure, and S vO2 is
ScvO2 as measured by the Swan, specifically.
In order to quantify the consistency of the sepsis model and success in producing
septic shock, some specific physiological endpoints were defined. Table 6.1 lists those
endpoints used for this set of experiments and the corresponding times it took, on
average, from perforation of the cecum to each endpoint (as well as the ranges across
experiments).
Pigs 4 and 5 required bolus doses of epinephrine (“epi”, a vasopressor) in addition
to the epi drip due to continually declining MAP. Other pressors used were dopamine
(Pig 2) and dobutamine (Pig 7). The relative effects of each are documented elsewhere 80.
The MAP was used to determine the epi drip rate, as well. Fluids were administered
147
according to the CVP. According to the Rivers protocol, the CVP should be around 12
mm Hg; however the exact endpoint for fluids is actually still in debate81. The CVP value
was lower in some experiments, especially in Pig 5.
Note that severe sepsis has
developed into septic shock when the patient is no longer fluid responsive.
Figure 6.9: MAP and CVP plotted with marks (+) indicating times at which pressors and
fluids were administered. The position of the marks on the y-axis is determined by
experiment; they have no actual numerical value.
148
Figure 6.9 again displays CVP and MAP, this time with the purpose of showing
the animals’ responses to fluids and pressors. The available MAP and CVP trends for
each experiment are plotted as lines; the marks (+) indicate times when pressors and
fluids were administered (or increased from the previous drip rate), respectively. Note
that the marks are placed on the y-axis only according to experiment. There is no real
numerical value assigned.
It is apparent from the plots that in most cases, fluids improved the CVP. Pig 7 is
a particularly good example of septic shock, as the animal is unresponsive to fluids after
about 4 hours. The pressors also proved effective in controlling the MAP.
b. ROx Data
The ROx data gathered in each experiment was analyzed using the process
described in Chapter 8. While this method has much room for improvement, it can be
useful for looking for general trends. The inability to aim the device at a vessel readily
prevented imaging at regular intervals corresponding to vital sign readings. However, all
images were time-stamped, so a general trend over time can be reported and compared
with other measurements.
Figure 6.10 shows progression over time of femoral SaO2 for all experiments for
which the data was available.
These blood draws were also not necessarily made
concurrently with the recording of other vital signs, but they usually match the timing of
ROx imaging. In addition to the development of sepsis, the inspired O2 was varied in an
unsuccessful attempt to calibrate the ROx. The timing and direction of the variations
(increased/decreased inspired O2) are displayed as +/- symbols in Figure 6.10.
149
Figure 6.10: Progression over time of SaO2 as measured via the CO-Ox for all
experiments with pertinent data recorded. Changes to the %O2 inspired are marked by
+/-, depending on whether the change was an increase or decrease. Again their positions
on the y-axis are determined by experiment.
For Pig 4, the changes in O2 inspired were relatively fast and large, ranging from
0-2.5 L/min of O2. This resulted in relatively wild SaO2 behavior in Pig 4; additionally,
the animal was one of the sickest of all the experiments. Even after the inspired O 2 was
maximized, the SaO2 continued to drop after about 4 hours. The inspired O2 variations
were more controlled in Pig 5, though they were also drastic. The superior health of Pig
5 throughout the sepsis experiment probably allowed it to sustain the variations better.
The variations in Pig 7 were the most controlled and most subtle, only decreasing the
inspired O2, which is evident by the relatively slow slope of the SaO2 fall-off.
150
Figure 6.11: Arterial measurements made on Pig 7 via the ROx (wavelength
corresponding to the minimum retinal arterial OD’s, corresponding to SraO2), CO-Ox
(%SaO2), and Pulse Ox (%SpO2). The error bars associated with the wavelengths are the
SEM between fit metrics as discussed in Chapter 8. Pig 7 was the only sepsis experiment
for which vessel identification was clear and a significant number of arteries were
analyzed.
Only the data collected from Pig 7 contained arteries that were readily
distinguishable from veins. The other data sets contain primarily images of veins. Figure
6.11 presents a comparison between SaO2 as measured by the CO-Ox, SpO2 as measured
by the pulse-ox, and the wavelength corresponding to the minimum in the retinal arterial
OD spectrum as measured by the ROx and analyzed via the method described in Chapter
8 (with subsequent error bars showing the SEM between the two fit metrics). Using the
BGO technique, the wavelength of the minimum OD is expected to correspond linearly to
SraO2. The ROx has not yet been sufficiently calibrated, though the ratio of ΔSO2 to Δλ
is roughly 3% per nm19, 20. Because the use of an erroneous calibration line would further
skew the data, only the wavelength (scale on the left) is used for general comparison to
accepted measures of SaO2 (scale on the right). Note that the vessel images from Pig 7
151
contained a significant glint, inhibiting the analysis and probably causing a significant
amount of the uncertainty and/or error.
The falling trend of the SraO2 nearly follows the trends of the CO-Ox and pulse-ox
measurements.
The dips around 3:30 can possibly be attributed to error; the large
variation over such a short time span further points to inconsistency in the data and/or
analysis process. Additionally, the low values around 1:10 could be similarly dismissed,
as the image for the point at 1:08 is not in best focus and the uncertainty in the point at
1:21 is so large. However, at that point in the experiment, pressors (both dobutamine and
an epi drip) were being added and increased, and the arterial line was removed and
replaced. Both of these events could have produced a drop in SaO2, though such a large
drop would be expected to show up elsewhere rather than just the eye.
This arterial data from Pig 7 is further discussed in Chapter 5 on calibration.
However, the majority of the retinal images were of veins, especially in Pigs 2-5.
Arterial SO2 is uniform throughout the body, making it preferable for calibration
purposes. During exsanguinations and reinfusion experiments on swine, SrvO2 trended
closely with ScvO270, 71; however, the relative behavior between SrvO2 and ScvO2 is not
well documented during sepsis, making the venous data less useful in instrument
calibration.
The progression over time of ScvO2 data from the femoral artery (measured via the
CO-Ox) is shown in Figure 6.12. Again, +/- symbols mark the changes in inspired
oxygen according to time and experiment (no values assigned).
This is to try to
distinguish changes due to sepsis from changes resulting from variation in inspired O2.
Compared to the arterial measurements, the venous response to change in inspire O2 is
152
more sporadic. Even though the response in the Pig 5 experiment is clearly visible in the
last half of the experiment, the venous saturation variations are not as dramatic their
arterial counterparts.
Figure 6.12: Progression over time of central venous SO2 (top) as measured via the COOx. Changes to the %O2 inspired are marked by +/-, depending on whether the change
was an increase or decrease. Again their positions on the y-axis are determined by
experiment.
As with arterial OD in Figure 6.11, the progression of wavelength corresponding
to the minimum in the venous OD spectrum is plotted in the top left of Figure 6.13. Note
the difference in x-axis range, because retinal data was collected from the second eye of
Pig 3, which required changing the pig to the prone position; however, this ruined the
calibration of the hemodynamic instruments, and no venous blood draws were made
during this extra time. The pig was moved instead of the ROx due to the positions of
other equipment at the time.
153
Figure 6.13: Top left: wavelength corresponding to the minimum retinal venous OD’s
(corresponding to SraO2) for all experiments. All other plots venous measurements made
via the ROx (wavelength corresponding to the minimum retinal arterial OD’s,
corresponding to SrvO2), CO-Ox (%ScvO2), and Swan (%SvO2) for each individual
experiment. The error bars associated with the wavelengths are the SEM between fit
metrics as discussed in Chapter 8.
Behavior of SrvO2 over the development period of sepsis is not immediately
generalizable via comparison between experiments (top left of Figure 6.13). The error
bars are the first quantifiable indication of sub-standard data quality. However, when
viewed by experiment, there is a more evident relationship between the venous saturation
measurements. The analysis method used on the ROx data seemed to work well enough
154
to see general trends, but it is sometimes difficult to distinguish between veins and
arteries; some of the points in Figure 6.13 might actually be data from arteries,
particularly in Pig 2.
No CO-Ox measurements were made for Pig 2, so only SvO2 is available for
comparison. Similarly, no ScvO2 measurements were available for Pig 7. The scales used
for the wavelengths of minimum OD’s vary from graph to graph with respect to the S vO2
scale; this is yet another indicator of the shortcomings of this set of ROx data and its
analysis. Ideally, the scales should be the same from experiment to experiment, with a
range of only about 35 nm corresponding to an SO2 range of 0-100%19, 20. Nonetheless,
the wavelengths of the OD minima follows the majority of the trends of S vO2
measurements and, to a lesser degree, the ScvO2 measurements (with Pig 4 as the
exception).
The Pig 2 experiment did not include changes to the inspired O2, so the drop in
MAP and SvO2 is most likely a result of septic shock. The ROx data points at 3:50 and
4:43 are the points which are most possibly arteries, which would explain the jump in
wavelength and relative SrO2. Another explanation could be that, in septic shock, the
retinal metabolism drops more quickly and takes longer to recover than what is seen with
the average over the whole body, which would result in an earlier drop in SrvO2 and a
later rise as compared with SvO2. However, the image quality and ability to analyze the
ROx data are both very limited in this experiment, so these would be very rough
conclusions at best.
As with the arterial ROx data for Pig 3, the time scale is different due to moving
the pig to image the second eye. The first eye imaged on Pig 3 had a blur caused by
155
corneal opacification that occurred when the eyelid was left slightly open after initial
anesthesia; only the point around 4:00 is from the first eye. The second eye imaged from
Pig 3 had a large glint, and often only four wavelengths were usable due to poor contrast
settings in the software. In order to image the second eye, Pig 3 was moved to the prone
position, which ruined the calibration of several of the hemodynamic monitors; no
venous blood draws were made after the shift, either.
The images from Pig 4, while not always in best focus, were better in terms of
staying on a single vessel through the majority of the experiment. However, due to
contrast settings, sporadic focus, and shortcomings in the analysis (particularly for 2
vessels very close together), this is one of the most uncertain data sets. The lowest points
with very large error bars in Figure 6.12 can probably be dismissed because of the large
uncertainty. It should be noted, though, that on the whole, Pig 4 developed septic shock
very quickly and severely; it had very erratic hemodynamic readings throughout the
experiment.
The image quality of the Pig 5 experiment was consistently the best of the first
five sepsis experiments. The uncertainty from the SO2 calculations is still too large to
make definitive conclusions, but it is the most complete comparison between S cvO2 and
SrvO2 of any of the sepsis experiments. The trend of the SrvO2 is continually downward
as of about 2.5 hours into the experiment, whereas the ScvO2 does not show any change
until the inspired O2 is lowered. This would support the hypothesis that sepsis is evident
in the SrvO2 before it can be seen in ScvO2 or SvO2; however the uncertainty of the first
SrvO2 point is so large, this is a rough conclusion. It is similarly reasonable to conclude
that the first point is erroneously high; if that point was actually the value at the lower
156
end of the error bar, a reasonable conclusion would be that the S rvO2 follows the trend of
SvO2. In short, a definitive conclusion cannot be drawn from this data set.
The image quality of Pig 7 was by far superior to those of any previous
experiment; they were generally well-focused, the resolution was about 100x better, and
the aiming was drastically improved; also, the paralysis of the eye made it additionally
easier to maintain focus in the same place for long periods of time. Unfortunately, the
improved focus emphasized the central glint. When the FSWF was changed in attempt to
improve the glint blockage, a different section of the eye was imaged, such that only an
artery is visible. Therefore, only about five images in the experiment contain analyzable
veins. They are timed close together toward the beginning of the experiment, so it is
difficult to draw many conclusions from the venous data from the ROx, though they do
seem to trend in the direction of the CO-Ox data.
There are several cases in which this relationship between SraO2 and SrvO2 has
been documented during hyperoxia. For instance, the retinal metabolism (related to the
difference between SrvO2 and SraO2) decreases as the inspired O2 increases61, 82. To our
knowledge, however, there are no examples of the relationship between SraO2 and SrvO2
have been documented during the development of sepsis. A comparison between the
ROx measurements of arteries and veins is displayed in Figure 6.14 in terms of the
oxygen extraction ratio, ER (Eq. from Ch. 2). Pig 5 was the only experiment in which
arterial and venous draws were made simultaneously, and Pig 7 was the only experiment
in which several images contained both an artery and a vein.
157
Figure 6.14: Comparison of oxygen extraction ratios ascertained from femoral vessels
(right, from Pig 5, using CO-Ox values) and from retinal vessels (left, from Pig 7, using
ROx images and Chapter 8 analysis methods).
The uncertainty in the retinal ER in Pig 7 is nontrivial, and it does not give a full
picture of the experiment, since all the points are from the original FSWF. Little can be
concluded from the calculated retinal ER. The femoral ER from Pig 5, however, is more
reliable, and the change in ER can be seen as it correlates with changes in inspired O2.
E. Discussion
Despite the preliminary nature of these sepsis experiments, there is some useful
information to be examined.
a. Sepsis Model
The sepsis model described here is, to our knowledge, the fastest controlled model
of sepsis to be presented, particularly in swine. In each experiment, the animal developed
severe sepsis and then septic shock. Additionally, each animal survived until the end of
the experiment, indicating success of EGDT in controlling sepsis. Once sepsis was
induced, the animals all remained critically ill, but full resuscitation is beyond the scope
158
of this set of experiments. MAP, CVP, and SvO2 were sufficient end points. HR, MAP,
SVR, and SvO2 all changed measurably.
Heart rate and temperature both increased fairly consistently throughout each
experiment, and stroke volume showed decrease toward the later hours of the experiment
as the bacterial toxin causes vasodilation. The CVP drops and cardiac output increases,
resulting in a drop in SVR. Depending on when measurement of SVR began, some
experiments show initial increase in SVR consistent with the body’s response to the
initial decrease in cardiac output as the sepsis sets in.
Fluids and pressors were used effectively to control CVP and MAP, respectively.
Toward the end of some experiments (e.g. Pig 7), the animal stopped responding to
fluids, indicating septic shock. The hypoxic dives affected the SvO2, but it was otherwise
a marker for needing epinephrine or dobutamine.
Other treatment was sometimes
necessary (such as electrical cardiac stimulation); the physician still has to respond to
each case individually, as different patients sometimes respond differently to the same
insult.
The ScvO2 sometimes trends differently than the SvO2, particularly in the later
measurements of Pigs 4 and 5. ScvO2 is the metric used by Rivers in EGDT68, which
showed a 16% improvement in mortality compared to standard care at the time. When
SvO2 was used as a primary metric in another study, the mortality rate was relatively
unaffected if not increased compared to standard care at the time83. For the purpose of
EGDT, either mixed or central venous SO2 can be used as endpoints in determining
treatment procedure where sepsis is involved66,
68, 84, 85
, though neither should be used
alone without other vital signs and endpoints83, 86. However, most studies dealing with
159
sepsis in humans use a catheter in the superior vena cava (at the heart) as opposed to the
femoral artery used in these pig experiments86,
87
.
The O2 consumption of the
gastrointestinal region increases in septic shock, particularly when the abdomen is the
source of infection88; the cerebral O2 consumption, by contrast, is maintained for a while
when the shock is setting in, affecting the SvO2 (as well as the SrvO2). This could be the
cause of the larger differences between SvO2 and SrvO2 in the pigs. Reinhart also noted
that the correlation between SvO2 and ScvO2 was the worst in the case of hypoxia
(Reinhart 1989). In general, however, studies point to the need for a combination of
relevant vital signs and end points for the success of EGDT89.
b. ROx Data
There is vast potential for the use of the ROx in sepsis and other shock
experiments, determining how the retinal SO2 responds as shock develops. However, due
to several factors (not the least of which is image quality and subsequent analysis), the
results from this set of experiments is limited to loose qualitative discussion in terms of
the physiological data.
The SrvO2 trends with the SvO2 and, to a slightly lesser degree, the ScvO2. The
retinal SO2 responds to changes in inspire O2, corresponding to results shown in previous
work20. There are, however, differences that cannot be confidently accounted for, since
the uncertainty in much of the data is large; it is not clear if it is error in the measurement
or a result of the physiology.
Ideally, an image would include both an analyzable artery and an analyzable vein.
From such an image, the cardiac output can be measured in addition to S raO2 and SrvO2.
While it is impossible to determine precise SO2 from ROx measurements without a
160
calibration line, the trend of SO2 in the retina can be compared with time and events
during the experiment, as well as with the trends of other hemodynamic measurements
made throughout the experiments.
The experience of the ROx operators affected the image quality, as well.
Targeting vessels and centering them in the image has become easier as aiming
mechanism improved, but there are multiple varying systematic factors for which the
operators must compensate throughout the experiment. The blank edges drift slightly, the
powers and wavelengths vary significantly, and the pig eye rolls, especially when the
SaO2 drops below 70%. All of these affect overall image quality and can be controlled by
the operator.
The widths of the blank edges of the images must be checked every time the ROx
is turned on, and they might have to be adjusted every couple of hours.
The edge widths
are easy to maintain via four potentiometer knobs. It is important that these edge widths
are approximately equal for all 5 images in order to ensure that the same region can be
analyzed in all wavelengths. It was a problem particularly in early experiments. More
attention was given to this aspect of the image in later experiments, fixing this issue.
To keep all 5 images as uniform as possible, the powers of individual wavelengths
had to be adjusted about every 30 minutes. The AOTF frequencies were checked when
the powers were adjusted, especially when it is evident that one of the wavelengths has
particularly low intensity despite a higher power setting. This became less of a problem
with the implementation of the PicoScope because of its greater effective bit depth. Until
then, the frame grabber was used with the contrast setting maximized to aid in targeting.
However, this made the system very sensitive to the power and PMT gain settings. Slight
161
power differences between wavelengths are very apparent in post-processing, often
ruining images because one or more wavelengths have saturated regions or are too dim to
be used. Another problem caused by the use of the maximum contrast setting in the
frame grabber software was vessel profiles that “bottomed out”. Since high contrast
amplifies brightness and darkness, near-zero pixel values in dark vessels would often be
assigned values of 0 in the frame grabber software. The true profile depth cannot be
determined with respect to the fundus, making it impossible to calculate the OD for that
wavelength. This is discussed further in Chapter 7.
Figure 6.15: Retinal image produced via frame grabber. The raw interlaced image was
the only image available to the ROx operator when the image was first acquired, and it
was roughly a third the size of the computer screen. This is an example from a pig eye in
vivo (Pig 5). Vessel profiles (blue lines) of the region selected in the bottom right image
(the sum of all 5images co-aligned). Compare to the images on the left. It is impossible
to calculate the OD values from any of the 3 wavelengths in the left column because their
minimum relative intensity value appears to be less than zero. Note that the red lines are
the vessel fits, and they are not good fits in any wavelength for in this case, probably due
in part to the presence of the glint and the lower resolution due to the orientation of the
selection. Even with the frame grabber, resolution is ~6x better in the rows than in the
columns.
162
For example, the vessels are too dark in Figure 6.15—the PMT gain needed to be
increased and/or the powers need to be equalized so wavelengths 1, 3, and 5 (left column)
are as bright as wavelengths 2 and 4 (right column). The contrast is set too high in the
frame grabber software. Figure 6.15 shows that this image is worthless for producing an
OD spectrum, as 3 of the 5 wavelengths have vessel profiles that “bottom out”, making it
impossible to calculate the OD of the vessel in those wavelengths. When the image was
acquired, the operator only had a larger version of the bottom right image in Figure 6.15
(roughly the size of about a third the size of the computer screen). It is not immediately
evident that there is any problem with the data set, judging by the raw interlaced image
alone. However, if the operator could have seen all 5 separate images (the other 5 images
in Figure 6.15) in a timely fashion, the problem would have been more evident. The
power and/or gain corrections could then be made and a second image could be taken.
Figure 6.17 shows the full resolution of the images, whereas Figure 6.16 shows
the lower resolution image produced for the ROx operator to determine image quality.
From this comparison, it is evident that the higher resolution improves image quality, but
for the sake of time, the lower resolution works for preliminary image quality control. It
is significantly more time efficient to use the low resolution images for the primary
quality check because the computer has to process edge straightening, wavelength
separation, and alignment just to show the basic images in Figure 6.16. The images in
Figure 6.17 include aligning and dividing out the reference, as well as binning/smoothing
each image.
163
Figure 6.16: Example of images the user sees within 20 seconds of snapping an image.
Note that the image on the left normally takes up the left half of the computer screen, and
the image on the right scales similarly. The raw interlace image (right) is still shown, but
the separated images are also shown (left). It is apparent from the separated images that
there is a blur in the vessel about 2/3 of the way down the image. It is consistent with a
corneal artifact seen with an ophthalmoscope, but it is not obvious in the raw interlaced
image. This example was taken from a pig eye in vivo (Pig 7).
Also, there are no contrast factors in play with the PicoScope, and the bit depth is
greater, which means that image acquisition is less sensitive to bright fundus and dark
vessels. With the frame grabber, the contrast and gain had to be balanced carefully. If
the contrast was not high enough, the vessel would appear washed out. If the contrast
was set too high, the fundus would appear saturated or the vessel values would “bottom
out” (i.e. be valued at zero; see Figure 6.15), since high contrast makes dark appear
darker and light appear lighter. The PMT does not output values of zero; even when the
PMT is covered, there is dark current producing some small amount of signal. Because
the PicoScope readout ultimately has greater bit depth and no contrast setting, vessel
profiles do not “bottom out”, so the operator has one less image quality issue to worry
164
about.
The greater bit depth gives more flexibility in power levels and uniformity
between wavelengths; a larger discrepancy can be present without degrading the image
quality.
Figure 6.17: Separated images from all 5 wavelengths at full resolution, after binning.
Note the improvement to the image quality (primarily due to noise reduction) as
compared to Figure 6.16. The improvement is not, however, necessary for a quick
qualitative analysis of the quality of the data.
Even with an improved aiming system and better image quality, it is not practical
to track a vessel when the eye rolls at low saturations. The paralysis of the eye in the Pig
7 experiment made a tremendous difference in the ability of the ROx operators to
maintain focus on a given region of the retina. This, paired with improved aiming and
better resolution, made for significant improvements in image quality, though the central
glint problem was further emphasized.
165
Poor image quality presented many issues that had to be addressed in postprocessing and analysis. Vessel identification within the image was a problem that had to
be addressed, as well as how to avoid the glint. As image quality improved, the ability to
select vessels drastically improved. The central glint is still a problem, and while the
analysis methods used attempts to deal with the glint via omission, it is still sometimes
too large for the automated analysis to handle.
It also difficult to determine the
effectiveness of other aspects of the analysis with the glint so significantly in the way.
For instance, selecting regions of avascular fundus used as reference for calculating
vessel OD is an issue that deserves further analysis, but with the current image sets, it is
difficult to tell what affect the selection of fundus has on the final OD calculation.
Though the glint may never be fully removed, minimization of the glint has high
potential for improving the ability to analyze ROx data consistently and automatically.
One challenge when analyzing several of these data sets is determining whether a
vessel is an artery or a vein. Ideally, this determination is made by identifying the optic
nerve head and finding an artery-vein pair.
The vessels can be distinguished by
comparison: veins are larger and usually darker, and if the pair intersects, the artery lies
beneath the vein. A vessel can be identified and traced further from the optic nerve head
for measurements in other regions of the eye. This is easier to accomplish with a large
field of view within the eye (such as with a fundus camera), but it has been successfully
carried out with the EOX-A, which had a similar retinal field of view to the ROx but with
significantly better maneuverability.
It was only in the last experiment (Pig 7, with implementation of the jack mount)
that the vessels are well identified. Because of the difficulty of aiming in the earlier
166
experiments, it was not feasible to trace vessels from the optic nerve head. The results
from the other experiments included in this section use the best judgment of the writer to
identify vessels. However, without pair comparison near the optic nerve head, it is
difficult to distinguish arteries from veins, especially when there is only one vessel in the
image. Having two vessels in an image does not always guarantee an artery-vein pair
because vessels fork and split. A smaller vessel may simply be an off-shoot of the larger
one (as was the case for most of the images from Pig 4, as determined by further
analysis).
Vessels can be identified in the blue-green wavelengths, but it is also helpful to be
able to view the vessels in the IR; arteries are nearly transparent, while veins are more
absorbing in the IR. However, the IR targeting was not available for any of the sepsis
experiments, either because it was not yet implemented or because of misalignment
during transportation to the animal lab.
F. Conclusions
The sepsis model demonstrated by this study is a fast and reliable way to induce
severe sepsis and septic shock in swine in a controlled fashion. Using this model, sepsis
can be induced and tracked as it develops into severe sepsis and septic shock.
The ROx was used in these experiments to collect retinal images in attempt to
track the retinal SO2 (especially SrvO2) through the progression of sepsis with hopes of
eventually using SrvO2 as an endpoint for EGDT. Due to problems with the central glint
from the vessels and general image quality, the only conclusion that can be drawn with
any confidence is that the SrvO2 and SraO2 trend, in general, with ScvO2 and SaO2,
respectively. Improvements have been made to the device to improve overall image
167
quality, but the central glint is still an issue that must be resolved before the analysis
methods can be further improved, in turn improving the quality of the results.
168
Chapter 7:
HUMAN EYE EXPERIMENTS
A. Introduction
The end goal of this work is to develop a retinal oximeter that can be used in a
clinical setting. This chapter describes the current capabilities of the ROx to that end.
Section B presents the considerations that are currently implemented for the sake of
human data acquisition. Section C describes the experimental setup and procedure for
obtaining images from a human subject. Section D contains results and data analysis, and
Section E gives conclusions.
B. Considerations for Human Data Acquisition
The ROx is designed to collect data from live human subjects, though the primary
goal is currently imaging pig eyes. Patient comfort is important to acquire repeatable
data while still maintaining device accuracy. The practicality and ease of use by a
clinician must also be taken into account. While the ROx-3 is not yet ready for clinical
use, some of these considerations are already being implemented. Many of these goals
and consideration have been drawn from the reports/design of the EOX and EOX-238, 58,
especially concerning convenience of aiming and speed of acquisition for a live human
volunteer.
a. Patient Comfort
The human eye is most sensitive to the blue-green wavelengths2. Therefore while
it is well within the ANSI safety standards to use the blue-green illumination for targeting
169
vessels, it is uncomfortable for the human subject, especially if his eye is dilated, either
by drugs or dark adaptation. The dilation drops were used in the past to acquire human
retinal images on the ROx47 and are even used in the most recent work for some retinal
oximeters like the Vesselmap90. For this reason, an IR laser, which is barely detectable to
the human eye, is used for targeting illumination instead. As mentioned in Chapter 4, the
IR wavelengths can also be used to distinguish between veins and arteries.
The IR wavelength must be within the detectable range of the same PMT used for
blue-green imaging, and the laser must be co-aligned with the AOTF-filtered beam from
the Ar++ laser. However, only one laser should be illuminating the eye at a time. The
custom electronics switch power between the AOTF and the IR laser driver, ensuring
only one is on at a time. The AOTF filters the Ar++ beam, controlling which wavelength
(if any) is passed into the system. However, even when the AOTF is off, some of the
light scattered in the AOTF crystal leaks into the system. While the amount that reaches
the eye is on the order of picowatts, it is still detectable to the human eye. To prevent any
discomfort or confusion to the subject, a shutter has been placed just after the AOTF to
block this leak. The shutter is controlled by the same signal that controls the AOTF, so
whenever the AOTF is on, the shutter is open, and whenever the AOTF is off, the shutter
is closed. As a further safety measure, the “off” position of the shutter is closed; in the
event of an electrical power failure, the shutter will close, preventing the subject from
exposure to the Ar++ beam.
In addition to use of the IR laser for targeting, the duration of the blue-green
exposure should be as short as possible (ideally, between 0.05 and 0.1 seconds). This is
for the comfort of the patient, but it is also to prevent the dark-adapted pupil from
170
constricting before the image is acquired. Pupil constriction reduces the amount of light
returning from the eye and potentially introduces vignetting if the nodal point of the
scanning illumination beam does not precisely match the nodal point of the eye. This is
necessary for the prototype: the aiming is not yet ideal, and the dark adaption of the eye is
crucial for image acquisition. In order to accomplish this, a DAQ has been implemented
to control the signals to the IR and AOTF drivers from the image acquisition software.
The DAQ is used to switch the path of the PMT signal between the frame grabber board
or the PicoScope, and it controls the shutter signal as well. Since these signals are
controlled from the software and ultimately the ROx computer, there is some dependence
on the timing of the computer processor itself. Sometimes the precision of the timing
between the AOTF turning on/shutter opening and the actual acquisition of the image
with the PicoScope is slightly out of sync. This results in occasionally missing a few
lines of the image, either at its beginning or its end. Compensation is therefore made in
the image display code, omitting blank sections.
When imaging in vivo, the subject should be stationary, with the head as steady as
possible. Whereas with the swine eyes, the animal is anesthetized with the eye sutchered
open, the ultimate objective of the ROx is to be able image the retina of a human subject
regardless of consciousness. For the purposes of testing this prototype, the goal is to
image the retina of an alert human subject sitting up. A head rest has been built to aid the
subject in stabilizing his head for image acquisition. The head rest is such that the chinto-forehead distance can be adjusted to fit the subject’s face, and its overall height can be
adjusted so that the patient is able to hold his head still relatively comfortably while he is
seated in a chair (Figure 7.1).
171
Once the head rest is adjusted for the subject, the operator can position the ROx to
illuminate the retina and begin targeting vessels. It is possible for the operator to aim the
ROx while the subject remains still. However, since the subject is conscious and alert,
the operator can ask the subject to move his eye rather than the ROx operator move the
entire device. This can be accomplished by giving the subject a target to focus on with
his free eye. The operator can control the position of the target; this is an effective way
of controlling the region of the retina at which the ROx is aimed since both eyes track
together.
Figure 7.1: Head rest for human patient, as seen from the point of view of the patient.
The horizontal white bars are the chin and forehead rests, and the height of the chin rest
can be adjusted via the silver knob on the right bar of the headrest. The entire assembly
can be raised and lowered electronically.
The human eye images had previously been acquired with the ROx before all of
these modifications were implemented47, but while the human subject was willing and
safe, the acquisition process was difficult for the ROx operators, very uncomfortable for
172
the subject. This has already been reported57, so this chapter focuses on the work since
those experiments.
b. Practicality for Clinical User
Several ROx design features are implemented with the ROx operator in mind.
Ultimately, the device needs to be usable by an operator with minimal specialized
training on the device. The operator should also have as much control as possible within
arm’s reach (or less), including focusing, aiming, and image acquisition.
The optical breadboard has been mounted on a jack that affords 5 degrees of
freedom of motion, as described in Chapter 4. Though the operator cannot control the
jack automatically, it is significantly easier to maneuver and aim than with the previous
set-up. The user can easily adjust the height, distance, and angle of the ROx once the
patient is comfortable and stationary.
There are two focusing lenses in the system. The first one (furthest from the
patient’s eye) can be moved back and forth along the optical axis to compensate for
myopia and hyperopia. For simplicity, it is mounted on a translation stage controlled by
an actuating micrometer. The driver for the actuator is located beneath the computer
monitor on the ROx electronics cart. The stage currently in use spans about 2 inches; a
larger range would be clinically useful for patients with more extreme defocus, but it is
sufficient for the current prototype.
Another consideration for the focusing lenses is their focal length. In the case of a
patient with perfect vision, the focusing lenses simply relay the collimated light, and the
depth of field is determined entirely by the effective focal length of the eye itself.
However, if the patient has some amount of defocus, the focal lengths of the focusing
173
lenses actually affect the depth of field when imaging the retina. Faster lenses, while
useful for minimizing the overall size of the instrument, decrease the depth of field and
consequently require a more precise positioning of the eye to achieve sufficient focus.
Originally, a pair of 30-mm-focal-length lenses was chosen in order to keep the device
compact on a single optical breadboard. These have been replaced with two achromats
with focal lengths of 60 mm. This literally give the operator more “wiggle room” when
focusing the device.
The amount of improvement varies, increasing as the lenses
compensate for greater defocus. As an aside, it also allows ±45 degrees of rotation about
a pig eye in vivo without the breadboard bumping into the operating table, which is
actually why 60 mm was chosen specifically.
In addition to considerations of ease of aiming and focusing, the ROx operator
should be able to evaluate the image quality in a timely fashion in order to repeat a
measurement if need be. Initially, when the frame grabber was the primary means of
acquiring images, the user would “snap” an image and the resulting raw interlaced image
would immediately be displayed. The user could then determine whether there were
errors or problems with the image, especially problems due to a skipped wavelength scan,
oversaturation of the image, or movement of the eye. However, since the interlaced
image was not yet divided into its 6 images by wavelength, it was difficult to make out
image quality and uniformity of power between wavelengths (see the discussion in
Chapter 6 for examples).
The PicoScope was originally implemented to improve resolution and pixel bit
depth. The data are initially presented as a large 1-D vector of pixel data and must be
sorted into 2-D form in post-processing, which is time consuming. By comparison, each
174
data set is ~100 times the size of the frame grabber image, which was already sorted into
a 2-D image. This consequently introduced the need to take speed of image presentation
into account. The USB port to which the PicoScope was connected was upgraded from
USB 1.0 to USB 3.0, reducing the time for the data set to reach the computer from
several minutes to a few seconds. The system is then limited by the write-time of the
computer processor. Previously, the frame grabber produced a single image file. The
PicoScope collects data from both the primary and reference PMTs, and the postprocessing for data analysis requires alignment between the two data sets, dividing out
the reference data, and binning the data before it is put into 2-D format and separated into
the 6 images by wavelength. This is currently a very time-consuming process (another
several minutes), which is unacceptable for clinical use.
Since the ROx operator only needs to perform a qualitative analysis of the image,
the resolution is not as critical; in fact, since the aiming takes place with the resolution of
the frame grabber, it seems reasonable that the user should be able to judge image quality
with the resolution of the frame grabber. In addition to saving the entire data strings from
the primary and reference detectors when an image is snapped, a second pair of data
strings is saved with 1/100th the resolution.
The small reference image is used to
determine line sizes to produce a square 2-D image, using the edge blanks. The small file
from the primary PMT is then used to generate both a 2-D raw interlaced file and 6
individual 2-D images by wavelength. This allows the operator to assess not only the raw
interlaced image, but the individual images by wavelength. Currently, the time from
snapping an image until the image is visually displayed takes 10-20 seconds using these
lower-resolution images.
175
Between the new optics jack mount, the focusing lens modifications and actuator,
and the addition of the PicoScope and the modifications to its code, the ROx has become
a more user-friendly device.
C. Experimental Setup
Over the course of a few years, the ROx has undergone updates and
developments. Several experiments involving imaging the human eye have occurred
during these developments, so the experimental setup was not always the same. In fact,
there was often either equipment malfunction or operator error. This section describes
the experimental setup for a working experiment.
The laser is turned on to warm up at 7.96 amps for an hour. Once the laser is
warmed up, the AOTF settings are adjusted so that the RF frequencies were optimal for
each wavelength, and the powers are equalized around 27 μW at the eye in all blue-green
wavelengths. While the power could safely be set to 55 μW, only 25-30 μW can be
reflected from the pellicle into the system in 457.9 nm, so all wavelengths are set to
match the maximum power available in 457.9 nm to keep the images relatively uniform
for a given PMT gain setting. This maximum power in 457.9 nm varies slightly, not only
every time the laser is turned on, but over time while the laser it on. The IR laser is
turned on and checked for alignment and focus with the blue-green. This can be done
with an Air Force resolution test target available in the lab, specifically using elements in
Group 3 (8-14 line-pairs per mm, which is the smallest group on the available target).
In addition to alignment and focus, the IR power must also be adjusted such that
the images appear equally bright in comparison to the blue-green images. Note that this
does not mean equalizing the powers, since the primary PMT does not have uniform
176
spectral sensitivity across blue-green and IR wavelengths (see Figure 7.2). This model
(Hamamatsu H10722-20 with a modified amplifier to increase bandwidth) has one of the
most uniform spectral sensitivities available across such a broad range of wavelengths.
Nonetheless, the IR sensitivity is only 60-70% that of the sensitivity to the blue-green
wavelengths, and the gain is nonlinear. This means that when the PMT gain is increased
by a given amount to brighten an image in the IR (e.g. during targeting), the same gain
change does not translate to an equivalent change in image brightness in the blue-green;
the blue-green image gets brighter than the IR image, even though the gain changes by
the same amount..
Figure 7.2: Spectral sensitivity of the primary PMT used in the ROx. The PMT is a
Hamamatsu H10722-20 model.
177
The current method of IR power adjustment starts with setting the blue-green
powers to ~27 μW (depending on the 457.9 nm power available). It has been empirically
demonstrated many times in both human and pig eyes that 20-30 μW is an acceptable
power range for acquiring useful retinal images. The brighter end of that range is
preferable because less gain is required, and therefore the images contain less noise.
Again, using experience and empirical evidence, the PMT gain is set such that, in a dark
room, the image of the square of illumination is just barely visible to the ROx operator
looking at the PMT output in video mode. White paper targets are too bright for use with
these powers, though other targets (such as a hand) can be used; if a hand can be clearly
imaged, the gain is sufficient. Once this gain is set, the illumination source is switched to
IR, and the power is adjusted so that the brightness of the image, either of the dark room
or the target, matches the brightness in the blue-green.
This process is not ideal; if the powers or gain are changed, another image
comparison between the IR and blue-green has to be made. If changes are made and the
comparison not checked, the result is often an acceptable or even excellent retinal image
in the IR, but when the image is snapped in the blue-green, it is either saturated or too
dim to see. This has been the cause of several failures to image human retinal vessels in
the blue-green.
Another consideration when switching between illumination sources is the
presence of ghosting or back-reflection from the focusing lenses. The lenses have an AR
coating for the blue green, but not for the IR. The effect is differing ghost reflections
between the two illumination sources. This has been another cause of experimental
failure; the reflection is reduced or removed in the IR, but the acquired blue-green image
178
is saturated with the reflection. Even if it does not cause experimental failure, this
reflection difference often causes difficulty. The focusing lenses have been positioned
such that the reflection is removed entirely, but when the IR is used for illumination, a
reflection appeared in the image such that it would impede targeting. This reflection
would be less critical if the IR was only used for vessel identification, since the bluegreen images are the key to measuring SO2. However, it is difficult for the operator to
determine best focus and target a vessel if a significant portion of the central field of view
is obscured by a reflection.
Once imaging settings (gain, power, alignment) are properly adjusted, the target
for directing the patient’s gaze is put in place. A flat mirror is placed such that when the
patient has his head placed in the head rest, the eye that is not being imaged can focus on
the ceiling. A red laser is shone on the ceiling within view of the patient as he looks into
the ROx: one eye is looking into the ROx, and one eye is looking at the red laser spot on
the ceiling. The ROx operator can then control the position of eye being imaged by the
ROx by moving the position of the red spot of light on the ceiling.
Once the patient is comfortably situated, the ROx aimed into the eye, and the
guide spot is properly positioned, imaging can begin. The IRB for the University of
Arizona has currently only approved one specific volunteer on whom this imaging
experiment can be performed, and there are no approved measures that can be taken to
change the SO2 of the volunteer. The purpose of this experiment is to simply determine
the data acquisition and imaging capability of the ROx when tested on a live, healthy
human volunteer.
179
D. Results
This chapter includes results from experiments after the PicoScope was
implemented. Dates are used to distinguish one experiment from the other; experiments
were performed at various stages of development. For the sake of comparison, the best
images collected with the ROx before the PicoScope was implemented are included in
Figure 7.3. The results were published in 201147.
a. March 26, 2014
Figure 7.3: Retinal images collected from the healthy volunteer using the ROx before the
PicoScope was implemented. Note the plot in the bottom right: it corresponds to data
from the vessel in the top right, showing that with careful analysis of an image by hand, a
reasonable OD spectrum can be determined and fit, producing a minimum that
corresponds reasonably well to the SvO2 of a healthy human.
The first experiment imaging a live human retina in the blue-green wavelengths
with the PicoScope installed took place March 26, 2014. Of several images that were
180
collected, only one was particularly useful (Figure 7.4). The resolution of the image is
markedly improved, and the reduction of noise is significant. However, there is still
uneven illumination, and there is image doubling a little less than half way down the
image. The angle of illumination is an aiming issue, but much of this is removed in postprocessing. The image doubling is due to lack of a trigger signal to the PicoScope.
In order to ensure that an entire congruous scan was achieved, a full period of the
slow-scan mirror (SSM) was collected via the PicoScope. This has been remedied; the
same signal that is used to trigger an image frame from the frame grabber is now split and
serves to trigger both the frame grabber and the PicoScope. This required modification
of the “CppSnap” code that drives the ROx. In order to grab 1 full scan (as opposed to 2
partial scans), the frequency and DC offset of the SSM driving signal, as well as the
software timing offset, must be properly adjusted.
Figure 7.4: Image of healthy human retina acquired March 26, 2014. Left: image
separated into its 5 wavelengths (and scaled blank, which is completely dark). Note the
difference in contrast in the 457.9 nm image. Right: all 5 images co-aligned and
summed. The rectangle indicates the region of interest for the profiles in Figure 7.5.
181
Figure 7.5: Vessel profiles in all 5 wavelengths and the OD spectrum resulting from the
automated fits.
However, note the difference in contrast between the 5 images. This is more
evident looking at the profiles (Figure 7.5).
The depth of the vessel profile is
significantly and unexpectedly shallower in the 457.9 nm image, resulting in a relatively
low OD and a large shift in the shape of the OD spectrum. This set of profiles indicates
that there is a significant amount of scattering or stray light in the 457.8 nm image that is
not present in the other 4 wavelengths. One of the strengths of the BGO technique is the
182
tolerance for power variations between wavelengths, especially with the addition of the
reference detector to divide out variations in laser power. However, the method is
crippled when scattering/stray light within the system is not spectrally neutral.
After investigation, it became evident that the non-neutral scattering spectrum
was being introduced at the pellicle. The reflection spectrum of the pellicle is neutralized
in the forward path by adjusting the powers of the each individual wavelength transmitted
by the AOTF.
However, the transmission spectrum is different because of the
compensation of power to neutralize the reflection spectrum, and the light initially
transmitted by the pellicle is very bright. Consequently, light scattered and transmitted by
the pellicle has a significant spectral dependence. The unscattered transmitted beam is
eaten in the system, but transmitted scattered light is detected by both PMT’s.
In order to rule out the primary PMT as a significant source of spectral error, the
background spectrum was measured at several different illumination powers, requiring
several variations of the PMT gain. The same spectral signature was present at all
powers, regardless of gain. For future spectral measurements, the illumination power
was set near 25 uW to best emulate a live eye experiment.
In order to reduce the effects of scattering from the pellicle, a small piece of metal
on a thin wire is used as a physical block. Because the return path is transmitted through
the pellicle at a different angle than the initially incident illumination beam, the physical
block is used to block the initial point of incidence out of the return path. Figure 7.6
illustrates the presence of the scattering from the pellicle and detected by the primary
PMT. The area shaded red contains scattered light from the beam initially incident on the
183
pellicle. When the physical block is properly in place, this scattering is blocked from the
view of the PMT.
The physical block was used in the first 6 pig experiments, but as the system was
improved and updated, the block was moved such that it was no longer functional. With
the PMT in place and no object in the view of the ROx, the block was placed such that
the PMT signal was minimized for all wavelengths, indicating minimum scattered light
from the pellicle.
Figure 7.6: Illustration of the light paths entering the field of view of the PMT. The blue
lines indicate the return path from the retina (or other object). The solid red lines show
the paths of the beam incident on the pellicle, and the dashed red lines roughly indicate
the cone of scattered light from the pellicle. The shaded region shows the cone of
scattered light that reaches the PMT unless it is blocked by the physical block.
Once this realignment was performed, the spectral signature of the system itself
was re-evaluated. A piece of white Spectralon was placed in the object plane with a
neutral density (ND) filter placed in an intermediate image plane covering a section of the
Spectralon. Ideally, the ratio of the relative intensities of the filtered region to the
184
unfiltered region should constant for all wavelengths. This test was performed before and
after the realignment of the pellicle block, and the results are shown in Figure 7.7.
Even with the block, the spectrum is imperfect. However with the block properly
in place, background images were acquired and subtracted from both the Spectralon
target images and retinal images from an enucleated pig eye on two separate occasions,
and subtraction of the background actually increased the noise, indicating that the
background is sufficiently low without being subtracted. Further investigation with better
testing is advisable for future work.
Figure 7.7: Normalized spectra measured by dividing ND-filtered Spectralon by
unfiltered Spectralon. Measurements were made before and after the block realignment.
b. May 8, 2014
With improvements made to the PicoScope triggering and the pellicle block
placed correctly, the healthy volunteer again allowed his eye to be imaged with the ROx.
185
While these images were the best ones collected with the ROx up to that point, they are
noisy and complicated to analyze. They are not in best focus, and several of the vessels
are horizontal (Figure 7.8).
The alignment between the IR and blue-green lasers, while very close, was not
close enough. When a glint is present, it is blurred out and difficult to see. There are also
reflections from dust on the focusing lenses. All these factors result in noisy vessel
profiles that are more difficult to fit.
Figure 7.8: Retinal image (summed over all wavelengths; the blank edges of each
wavelength were not illuminated evenly, resulting in darkened edges in the summed
image) in healthy human. Region of interest designated as rectangle on image at the top
right. Note the visibility of the optic disk in the top right of the image. The bright spot in
the center left is a reflection from the one of the focusing lenses. An example of the
automated analysis of a horizontal vessel is shown to the left, with resulting OD plot
shown below. Note the resolution of the profiles.
186
The automated data analysis had not yet been automated to deal with horizontal
vessels, and this data set forced the issue. The automated program was improved to
analyze horizontal vessels; however, the consequence of imaging a vessel in the
horizontal orientation is the decreased resolution of the profile. This results in less
reliable fits to the profile. Figure 7.8 shows an example of the automated analysis of a
horizontal image. This is one of the better fits, both of the vessel profiles (shown in the
figure) and the parabolic OD spectrum, and it is still sub-standard.
Figure 7.9: Image of retinal vasculature in a healthy human eye. Note the region of
interest selected contains an artery-vein pair, and consider that when using a rectangular
region of interest, it would be difficult to select a significant region with only one of the
vessels. However, an addition to the automated program allows the user to select only
one of the vessels. An example of the resulting profile is shown on the right, where the
user has selected the vein (darker vessel on the left). At this point in the experiment, the
SSM signal is not perfectly synchronized with the data acquisition, so there is some
image doubling at the bottom of the image. This image is the sum of all 5 wavelengths,
but the edge blanks were not evenly aligned, resulting in the darkened edges in the
summed image.
Another required addition to the automated analysis included more careful
selection of a vessel when dealing with an image with two parallel vessels close together
(Figure 7.9). This was the first data set collected with the ROx that required such
187
precision in selection; simply looking for differentials in a profile is no longer an option.
The current automated process requires one extra step by the user: selecting which vessel
to use. This is relatively simple, since the user has already straightened the image
according to the vessel of interest; however, it is a necessary step to omit other vascular
regions in order to properly normalize the vessel profiles to the avascular fundus. These
additions to the code are explained in further detail in Chapter 8.
Figure 7.10: Images of a human retina near the optic nerve head (summed over all 5
wavelengths, with darkened edges due to poor alignment of the edge blanking). The
rectangles indicate the selected region of interest (ROI). Note that a vein-artery pair is
present in both ROIs. The white spots are reflections from the focusing lenses, and the
SSM timing signal is not perfectly in sync with the software, resulting in slight doubling
at the bottom of the image. There is vignetting in the image on the right, but both of
these images are in decent focus. In fact, the glint is visible in the artery (bottom of the
pair) in the image on the left.
Not all of the images were so poorly focused (see Figure 7.10), but the horizontal
nature of the vessels requires that the automated analysis program be improved further
and/or the image quality be improved. The lenses should be cleaner, and vignetting
should be better avoided.
188
c. June 20, 2014
In previous work on the EOX by Denninghoff et. al, the pinhole at the image
plane of a 2.2X system was 200 μm in diameter58. The illumination spot on the retina
was 40-60 μm in diameter, so the 200 μm pinhole was about 150% of the image of the
spot size. In the EOX, however, the central glint was addressed via crossed polarizers.
The function of the FSWF in the ROx is to block the specular reflection from the
illumination spot on the retina and collect the scattered light that has interacted with the
illuminated region.
Originally, the magnification from the retina to the image plane at the PMT was
calculated to be 2.4X, such that the 2 mm pinhole of the FSWF accepts light from an 820
μm spot on the retina, of which the central 100 μm is blocked by a 240 μm wire bisecting
the pinhole57. This effectively blocks the illumination spot itself (and the resulting
specular reflection) and allows the PMT to detect the scattered light. However, after
rearranging the optics such that they fit on a single bread board, the new magnification is
closer to 6. This means that the wire was not effectively blocking the glint because it was
~3 times too small. Similarly, the pinhole was more tightly confocal than necessary,
accepting scattering from a region ~330 μm in diameter—almost the same diameter as
large vessels. This significantly inhibited the collection of light scattered from blood to
fundus surrounding the vessel.
While the system magnification was under further analysis, an ad hoc replacement
was assembled and applied to the system, and the human volunteer sat through a brief
imaging session. The experimental FSWF was a 5/16" hole (7.9 mm) bisected by a 36
gauge wire (127 μm) suspending a 1.4 mm round metal block. The diameter of the
189
collection area on the retina, consequently, was around 1.3 mm, with about 230-μmdiameter spot blocked from the center of it to minimize the specular reflection.
Figure 7.11: De-interlaced images from human retina collected with the 7.9 mm pinhole.
The artery is the lower vessel, and the vein is the upper vessel, evidenced by the slight
difference in color (veins are darker in the blue-green, especially at 514.5 nm). The white
spots are reflections from the focusing lenses, and the white vertical line is dust on the
SSM (which moves vertically with respect to the image).
The resulting images, while in fair focus, have very low contrast (example in
Figure 7.11). In fact, at first glance, the scattering problem in the 457.9 nm image seems
to have returned. However, the vessel fits (via automated analysis, Figure 7.12) produce
190
a reasonable OD spectrum, with a 6-nm wavelength difference in the respective minima
for the artery and vein (bottom and top vessel), respectively. Even so, the shallowness of
the vessel profiles actually make it very difficult to determine the vessel fit in the first
place. It is non-trivial to attempt by hand, and the automated analysis did not produce
sufficiently consistent or reliable fits. There appears to be a glint visible on the artery,
which may also indicate imperfect placement of the FSWF.
Figure 7.12: Vessel profiles for the artery (left) and vein (right). Due to the shallow
nature of the profiles, the reliability of the vessel fits is questionable. However, the
parabolic fits to the OD spectra have minima at 490.3941 nm and 484.5519 nm for the
artery and vein, respectively, indicating an approximate difference in SO2 of 18%.
These images (Figure 7.11) have very low contrast. The experimental FSWF,
while not totally ineffective for imaging, did block too much of the light from the central
return path. The light nearest the vessel as it was illuminated—light that had interacted
with the blood—is the most desirable signal, but it is also nearest the specular reflection.
The spot blocked by the FSWF was 3-5 times the size of the illumination spot on the
retina (and potential vessel glint). The illumination spot is larger at the edges of the
191
image due to distortion. The central block must be large enough to block at least the
majority of the glint, but small enough that it allows for detection of scattered light
nearest the illumination spot.
The pinhole was large enough, though, to collect at least some of the useful
scattered light undergoing blood interaction. This is why there is any parabolic shape to
the OD spectrum. However, a significant amount of the light collected had not interacted
with the blood (such as light from the glint that has been scattered and reflected about the
avascular fundus), resulting in the shallow vessel profiles. The pinhole must be large
enough to collect reasonably measureable signal in spite of some central block
obstructing the specular reflection from the vessel.
Despite the overall failure of this FSWF, the robust nature of the BGO technique
is displayed. There is about 6 nm difference in the minima of the artery and vein (where
the artery minimum is greater). For a healthy human whose S aO2 has been measured
around 97% with an off-the-shelf pulse oximeter, this would roughly indicate an SrvO2 of
about 79%, using the slope of the intravitreal illumination calibration line of -3% SO2 per
nm.
d. June 23, 2014
Once the system magnification was better determined to be 6X, a set of 6 FSWFs
were constructed. They were tested for contrast and resolution by comparing images of
the Air Force target, Group 2. The details of this test and a better subsequent test are
discussed in Chapter 10. The FSWF chosen had a 4 mm pinhole and an 800 μm bisecting
wire to pass light from a region of the retina approximately 670 μm in diameter while
192
blocking the central region approximately 130 μm across. With the new FSWF in place,
the eye of the human volunteer was imaged again.
A few decent images were acquired, but the focal alignment between the IR and
the Ar++ lasers was not sufficiently close. During targeting using the IR, the quality of the
images seen by the ROx operator was clear and focused, similar to the quality seen later
in Pig 7. However, the quality of the images obtained in the blue-green was out of focus
(Figure 7.13). This makes it difficult to observe and analyze any glint present on the
vessel, as it is possibly blurred out. This set of images is therefore not what is needed to
determine the effectiveness of the FSWF.
Figure 7.13: Image and corresponding vessel profiles (the selected region is boxed off)
from a healthy human eye. The resulting OD spectrum is on the bottom right. The image
is the sum of all 5 blue-green wavelengths. The IR targeting image was in best focus
when this image was acquired. The minimum value of the parabolic fit to the OD
spectrum is 481.6067 nm.
193
Figure 7.14: Image and corresponding vessel profiles (the selected region is boxed off)
from a healthy human eye. The resulting OD spectrum is on the bottom right. The image
is the sum of all 5 blue-green wavelengths. The IR targeting image was purposefully
defocused when this image was acquired. The minimum value of the parabolic fit to the
OD spectrum is 490.7706 nm.
Attempts were made to find blue-green focus relative to IR focus, but it was
quickly determined that the better solution is to improve the relative focus. For example,
Figure 7.13 was taken once the IR targeting image was deemed in best focus by the ROx
operator. Figure 7.14 is an image of the same vessel, but the ROx operator purposefully
defocused the image while targeting in the IR by moving the focusing lens via the
actuator. Similarly, a third image of the same vessel was acquired, but the operator
defocused the image, this time moving the focusing lens approximately the same distance
in the opposite direction. The vessel is not visible at all in this third image. Note that the
focus is similar in the first two images (Figure 7.13 and Figure 7.14).
194
The profiles are relatively noisy compared to those of pig eyes. This could be
because the vessel is smaller, so its size in the profiles is small compared to the noise.
The noise is, however, also affected by the size of the pupil; dilated pig eyes have much
larger pupils than the approximately 7-mm-diameter dark adapted human pupils. This
means a reduced amount of light in the return path, so the PMT gain is increased, thus
decreasing signal-to-noise ratio. The SNR of the fundus pixels for Figure 7.13 is 416.7
compared to an SNR of 714.3 in the fundus of an enucleated pig eye. On the respective
vessels, SNR is 86.4 in the human eye and 100.0 in the enucleated pig eye.
The profiles are taken from the regions of interest selected by the user (indicated
by the boxes on the images). Only the parabolic fit metric was used to fit the profiles and
perform the automated analysis (details in Chapter 8). The profile fits from the vessel fit
metric were very poor and entirely unbelievable. Those profiles (particularly in 457.9
nm) do not seem to properly correspond to the OD spectra, but this could be because of
the factor from the “spectrally neutral fundus” that is applied after the vessel profile is
determined (Ch. 8). However, it could also be indicative of improper placement of the
pellicle block again.
Both analyses from Figure 7.13 and Figure 7.14 produce reasonable OD spectra,
but the resulting SO2 values have a difference of about 30%.
This is completely
unreasonable for data from the same vessel collected within minutes of each other,
especially when there are no significant changes in the physiology of the patient. The
inconsistency of this result reiterates the need for better image quality (including better
aiming/targeting ability), and improvement to the automated analysis.
195
A further example of the limitations of this image set is the difference in
wavelength corresponding to the minimum OD between the two images is large. The 9
nm difference corresponds to an SO2 difference of approximately 27%. Images of the
same vessel taken within a minute of one another should not have a difference in minima
of more than 1 nm. This result indicates a shortcoming in image quality and the inability
of the analysis code to compensate, as well as the persistent nature of this problem.
Figure 7.15: Image from a healthy human eye (sum of all 5 blue-green wavelengths).
This image was a serendipitous image acquisition because of the quality of focus, but it
was not in best focus in the IR and not readily repeatable.
Figure 7.15 was acquired during the same experiment, though it was not readily
repeatable, despite efforts to do so. The quality of focus in this image indicates that the
system is capable of acquiring better images if the targeting alignment was improved.
The contrast is also visibly improved from the previous experiment, indicating that the
change in the FSWF is at least in the right direction. However, the profiles from this
image do not produce results similar to those of Figure 7.13 and Figure 7.14, considering
that they are probably images of an artery and the vessel selected in Figure 7.15 is a vein.
196
Additionally, this image is interesting because of the reflection from the vessel on the
right. One possible explanation is incorrect positioning of the filter.
e. June 25, 2014
Before performing this experiment, the IR was refocused to match the blue-green
illumination path using the Air Force target, Group 3. The pellicle block alignment was
checked, and the FSWF replaced with the original FSWF (2-mm pinhole, 270 μm
bisecting wire) used in previous experiments; this was done in order to further increase
contrast. There was a misunderstanding as to the actual goal of this experiment: it should
have been only to obtain better data from a human eye. Instead, the goal was understood
to be confirmation of instrument readiness before a live pig experiment. This was a
mistake in that the larger FSWF used in the previous experiment should have been used
in the actual pig experiment.
With final confirmation as the goal in mind, this experiment should have been
done with an enucleated pig eye, but none were available until after the experiment
(discussed in Chapter 10). Due to time limitations, the more immediate goal of this
experiment was to collect analyzable blue-green images from a human eye using IR
targeting after changing the FSWF.
The human volunteer participated in two imaging sessions for this experiment, the
first of which was a failure due to power differences between the IR and blue-green. The
IR was too dim compared to the blue-green; the operator was able to target and see crisp,
clear images in the IR, but the blue-green image acquisition resulted in washed out,
saturated images because the gain was too high. This was discovered and fixed quickly,
and the human volunteer sat for another round of imaging. One decent image was
197
collected (Figure 7.16). However, there was a back reflection from the focusing lenses in
the IR, making targeting difficult. To compensate, the ROx operator slightly adjusted the
offending lens. This made targeting easier, presenting the operator sharp retinal images,
but the subsequent blue-green images contained no vessels or definition. It was later
determined that this was due to back reflection in the blue-green washing out the images.
The re-adjustment of the lens had introduced the reflection; the difference in back
reflection between the IR and blue-green is problematic. In fact, any presence of back
reflection presents complications in image acquisition.
Figure 7.16: Images of human retinal vessels (left) and the corresponding vessel profiles
(right). The profiles are those of the darker vessel, the selection for which is shown in the
bottom right image (all 5 wavelengths summed). The vessel profile fits (in red)
correspond to the OD spectral plot at the bottom right, with a spectral minimum at
476.0564 nm.
The contrast in Figure 7.16 is greater in the 514.5 and 457.9 nm images than the
488.0 nm image, and the vessels are relatively clear and well-focused. There is some
198
noisiness in the images, probably due to higher PMT gain. However, upon further
analysis, the automation did not handle the profile fits well enough (the best is shown
here). The OD spectrum, while relatively parabolic in shape, produces a minimum
wavelength of 476 nm, which would correspond to a SO2 value around 20% considering
the intravitreal illumination calibration line20. This could be a result of placement of the
FSWF, insufficiencies in the analysis process, stray light within the system, etc.
However, without more images or a calibration line with which to compare, there is not
much that can be said with confidence about this analysis.
E. Conclusions
There are several purposes for this line of experiments, many of which were
successfully met. The ROx was upgraded to acquire higher-resolution retinal images
from a live human volunteer using only dark adaption of the eye. The volunteer is able to
sit in a more comfortable and stationary position with the addition of the head rest. The
ROx operator can direct where the volunteer is looking, allowing the operator to target
regions of the retina.
The precision and consistency needed to acquire dependable SO2 values from a
human eye are lacking. While the IR targeting works, it presents inconsistencies due to
spectral sensitivities of the primary PMT and the AR coatings of the focusing lenses.
There are issues of co-alignment and matching focus with the blue-green illumination
path. The focusing issue may have been resolved in the last experiment, but it could not
be well-verified because of the reflections and power differences.
The reflection issue could possibly be resolved by replacing the 2 focusing lenses
with a pair of OAP mirrors coupled with a pair of flat mirrors, allowing for similar
199
motion. This would introduce some aberrations off-axis, but the lenses are already being
tilted and/or translated to overcome these back reflections, which also introduces
aberrations. It has the additional benefit of allowing for longer focal length mirrors
without lengthening the size of the system. This would further improve the depth of field
for defocused eyes. A more detailed suggestion is included in Chapter 10.
The spectral sensitivity of the primary PMT is a nontrivial issue with experiment
setup, particularly between IR and blue-green images. With the current system, the
powers of the IR laser and the AOTF, as well as the primary PMT gain, must be properly
determined such that the primary PMT produces images of similar brightness in both
spectral ranges.
The quality of human retinal images obtained from this device is not yet on the
level of other devices, such as the EOX-2 and others38, 40, 91. However, the most pressing
goals of the instrument do not yet include the ability to obtain dependable SO2 values
from human eye; the more pressing matters are the calibration line (which cannot be done
on humans for safety reasons), glint removal, and the automated analysis program. Once
the automated analysis can be shown to be consistent and dependable and a calibration
line has been established for the ROx, human retinal imaging will increase in priority.
200
Chapter 8:
DATA ACQUISITION AND ANALYSIS METHODS
A. Introduction
The ROx-3 is a retinal oximeter that utilizes the blue-green oximetry technique.
This chapter describes the image analysis required to produce data for which BGO can be
used. Section B describes the data acquisition process and preparation of raw data into
analyzable images. Section C details vessel and spectrally neutral fundus identification
methods. Section D gives account of methods used for obtaining spectral data from the
images, including the vessel fit equation. Finally use of the BGO technique is detailed in
Section E.
B. Image Preparation
The images initially acquired with the ROx were two-dimensional interlaced
images supplied by the frame grabber board. To reduce overall noise in the images, a
PicoScope high-speed analog-to-digital converter (ADC) was later incorporated in the
system. The ADC produces a single line of data with higher resolution than the frame
grabber. With the PicoScope in place, the frame grabber is only used for real-time
streaming for targeting purposes and no longer for image acquisition (for the sake of
time). The image preparation for both cases is discussed in this section.
a. Frame Grabber Image Acquisition and Preparation
The Foresight Imaging I-50 HSN frame grabber used in this generation of the
ROx is currently considered to be one of Foresight’s “legacy” products, and the
201
associated software development kit (SDK) and example programs are functional albeit
archaic. Based on the available examples, the image acquisition is controlled by a C++
MFC API created in Visual Studio 2010. It was updated and edited to allow the user to
“Stream” images and to “Snap” images fed to the frame grabber by the PMT.
The
program is entitled “CppSnap”, shown in Figure 8.1.
Figure 8.1: CppSnap GUI program used for image acquisition. This is a snap shot
example of “Streaming” mode, using the IR targeting laser to image 2 retinal vessels in a
live swine.
Streaming mode is used for targeting, wherein the data is displayed as a real-time
streaming video in 400x780 pixels, since this is a convenient size on the computer screen.
When an image is snapped, a 400x780-pixel interlaced image is instantaneously grabbed
and saved to a unique time-stamped .png image file with a bit depth of 10 bits. As soon
as the file is saved, the program goes back to streaming data. This file can then be read
into Matlab for further manipulation. It has been shown that it is possible to run a Matlab
202
analysis of the image from within the C++ program, but it requires significantly more
computing power than the older basic computer can offer in order to be practical.
The “brightness” (bit value amplitude) and “contrast” (scale factor) can be
adjusted in the graphic user interface (GUI) for the frame grabber under “Options” (see
Figure 8.1), but an incorrect setting can be detrimental to the data. In many of the early
images, these settings were not properly adjusted. In some, the bit value could not reach
its true minimum because the brightness setting imposed a significant offset. In others,
pixel values (particularly on vessels in an image) are zero because the contrast setting
was pushing low values to zero. Similarly, still others show saturated fundus too readily
due to contrast settings pushing high values to the maximum.
Once in Matlab, the interlaced image is separated by wavelength. Occasionally a
single row of a wavelength will be skipped (probably due to a minor glitch in the AOTF
signal electronics), throwing off the image separation. Compensation has been made to
this end, basically removing any segment of the image that has less than 5 rows between
blank rows. Then the image is divided into 6 individual images by row: one for each
wavelength and one blank. Rows 1, 7, 13, etc. go into one image; rows 2, 8, 14, etc. go
into the next image, and so on. The images are then ordered according to wavelength
using the blank image (with the lowest mean pixel value) as the reference point. “1”
corresponds to 514.5 nm and “6” corresponds to the blank. The even images are then
flipped horizontally so that all the images have the same orientation.
203
Figure 8.2: Images separated by wavelength. Systematic noise is sometimes present in 1
or more wavelengths (see wavelength 3 here) and must be filtered out. The prompt
allows the user to select which wavelength (if any) to filter.
204
Figure 8.3: If a wavelength has been selected for filtering, a couple of rows are shown in
frequency space (symmetrical about 0). The user is prompted to draw a box around one
set of unwanted frequencies.
In the first sets of pig experiments, there were instances of systematic noise within
a single wavelength (probably a problem with the customized AOTF switching circuit).
This should be removed in a fast and relatively consistent manner. All six images are
displayed and labeled. The user is then prompted to select the wavelength with the
systematic noise (since so far only one wavelength has ever been an issue at a time), or
indicate that all wavelengths are fine (see Figure 8.2). If a wavelength is selected, the
Fourier transforms (FT) of a couple of rows in that image are taken using the fast Fourier
transform (fft) function in Matlab and summed. Summing the FTs helps to highlight the
frequencies that are persistent throughout the image. The summed FTs are displayed, and
205
the user is prompted to select a region of frequencies to be removed via a cosine filter.
Due to the symmetry of a cosine filter, the user selects a single region, and its
symmetrically corresponding region is automatically selected as well. For every row of
the image (in the selected wavelength), the fft is taken, the selected frequencies are
removed, and the row is then transformed back into its original space domain. The
filtered image is then displayed for the user to check. The process can be repeated if
necessary.
Figure 8.4: A user-selected region is shown; ideally the amplitude would fall off with
distance from 0. The box has been drawn around an anomalous peak.
206
Figure 8.5: The 488.0 nm image before filtering (left) and after filtering (right). The
horizontal flip has no effect on the rest of the analysis.
b. PicoScope Data Acquisition
When the PicoScope was incorporated into the ROx, the same MFC-type C++
API program was used for image acquisition. Examples provided with the PicoScope
SDK were then integrated into the program. Two PMTs are used: the primary PMT
detecting return light from the eye (corresponding to Channel A) and the reference PMT
detecting the beam transmitted by the pellicle on the first pass and scattered off of
Spectralon with 2% reflectivity (corresponding to Channel B). The Foresight frame
grabber is still used for aiming and targeting purposes, but now when an image is
“snapped”, two unique time-stamped binary files are created containing the data as 2 1-D
vectors of data with a bit depth of 8. One file (“A”) contains data from the eye and one
file (“B”) contains the corresponding data from the reference detector.
In addition to the input from both PMTs, the PicoScope has an external input port
(EXT) which can be used for triggering. In order to produce 2-D images in real time, the
frame grabber uses a modified version of the signal that controls the slow scan mirror
(SSM) to initiate an image frame. This same signal is sent to the PicoScope EXT port
207
and used to trigger data acquisition when an image is “snapped”. When the user clicks
“Snap”, the program waits for the next rising edge of the modified SSM signal, initiates a
programmed delay before acquiring a data for half of the period of the SSM. This
ensures that a complete single image is acquired (no image doubling), and the
programmed delay is such that the acquired image matches the targeting images produced
via the frame grabber.
The Foresight frame grabber and the PicoScope cannot obtain an image
simultaneously as that would require splitting the voltage output from the PMT.
Consequently, a new method is required to preview the “snapped” image acquired by the
PicoScope. Because time is of the essence and maximum resolution is not necessary for
a preliminary view of the acquired image, a 1/100th scale sub-sampled set of the data
from both Channel A and Channel B are saved as additional time-stamped binary files.
They are labeled with a suffix of “Sm” and “SmB” (as opposed to “A” or “B”) to
differentiate them as the “small” files. They are read into a custom Matlab function
designed to display the data in image format. The process is explained later in this
chapter.
The PicoScope can be treated like an oscilloscope in that the user can control the
voltage range and offset of the data as well as the sampling rate (“time interval”). For the
current system, the range is set to ±2V with an offset of 2V for both PMTs. Using
different settings for the two PMTs seems to introduce discrepancies between the data
sets that are not conducive to dividing one image by the other without problematically
propagating those discrepancies. The time interval for both channels is minimized for
best resolution: 500 MSa/s. The timing synchronization of these two files has been
208
confirmed; any small difference in timing is removed during post-processing. The data
from the PicoScope is a string of data values stored as a 1-D vector in binary, or .bin,
files to conserve space (as opposed to text files), since it saves each 25 Mpixel data set
(collected in 0.05 s) at about 50 MB in size. Note that the pixel values can be negative
due to the 2V offset (which does not translate into the pixel values). To compensate, the
difference between zero and the minimum pixel value (overall, between both reference
and primary data) is added to all pixel values. To prevent any division by 0, all values
are additionally increased by 1. Values for pixels illuminated by the laser are on the
order of 104, so this should be negligible.
As mentioned in Chapter 4, a DAQ has been added to the system. This includes
addition of the DAQ driver into the program code. Again, a software development kit
(SDK) was helpful in this undertaking. The program is set such that upon initiation, the
DAQ sends a 0V signal to the analog switch, directing the primary PMT signal to the
frame grabber where it stays for the duration of “streaming” or targeting. Upon clicking
“Snap”, the DAQ sends a +5V signal to the analog switch, redirecting the primary PMT
signal to the PicoScope (which is already constantly receiving the reference data). When
the data acquisition is complete, the DAQ returns the PMT signal to the frame grabber (at
which time the real-time data streaming can resume). A hard switch also allows the
option for the user or the DAQ to control which laser is in use. The same DAQ signal
can switch illumination sources: the IR is on when the DAQ signal is 0V, and the AOTF
is on when the DAQ signal is +5V. Because of this illumination control, it is important
to keep the 5V pulse as short as possible while maintaining illumination and open
shutters during the PicoScope acquisition.
209
The command sequence for data acquisition using the PicoScope can be
summarized as follows:
1. Open API executable program.
2. “Stream”, using the Foresight frame grabber to display real-time images
acquired by the ROx for targeting purposes.
3. “Snap” an image using the PicoScope, saving 4 time-stamped data sets
(including the 2 small subsets).
4. The small primary subsets are automatically manipulated to produce one
“raw” image and six wavelength-separated images.
5. Streaming can be resumed and the process repeated as needed.
It should be noted that many of the image display controls associated with the
frame grabber are still functional during streaming/targeting. However they have no
affect on the data acquired from the PicoScope. The PicoScope is used during the
initialization of the ROx to compare primary and reference PMT outputs, as well. The
reference signal should be similar in scale to the signal from the primary PMT.
c. PicoScope Image Preparation from Sub-sampled Data
It is currently critical for the user to be able to view the image shortly after it is
snapped in order to perform a brief qualitative analysis. The original 48-MB .bin files
contain data for 25 Mpixels, which is too large for the computers currently available to
this group to process in a time-efficient manner. Therefore, before any further image
preparation is attempted, the “Small” data files are opened using a custom Matlab
function. Blank edges and blank lines from the reference are used to determine the width
210
of the image (the number of pixels per line in the fast-scan direction). The reference data
set is used to avoid identification uncertainty between blank edges and vessels.
Unlike the Foresight frame grabber, the PicoScope 3207B in the current system is
not triggered by the fast scan mirror. A block of data begins in an arbitrary position in a
random line. The data is first handled in its 1-D vector form. Because the data often
begins mid-fast scan, the first extreme rise in pixel value is used to define the beginning
(top left corner) of the image. Because of the possibility for power variations between
wavelengths, only the maximum over a short range is considered. It is best to include the
brightest wavelength to avoid confusion in the code, so the first 3600 pixels (roughly 6
line widths) are included in finding a maximum. When the difference over a 15-pixel
range becomes greater than 30% of the maximum over this region, the position of this
first threshold differential (edge) is recorded.
Once the starting point is determined, an approximate line size must be estimated.
From past experience, the line size of total data sets (A or B data) has consistently been
on the order of 63,500 pixels. The line size of the small images is therefore roughly 635
pixels. The next two differentials that meet the threshold value are found in a similar
manner as the first. To keep from finding multiple differentials above the threshold for
the same line, the data pointer is advanced 300 pixels before the search for the next edge
begins. The distances between the first and second edges and the second and third edges
are found and compared. The minimum distance is used as the initial approximate line
size, as this ensures there is not a blank line factoring into the estimate. This was also
used to determine where the image should begin (again avoiding starting with blank
lines). The data is truncated such that an edge begins 50 pixels after the beginning of the
211
image; the 50-pixel buffer is to ensure the blanked edges are included on both sides of the
image.
The data is then analyzed pixel by pixel, creating a “mask”, or binary data set that
identifies any pixel with a differential of 10% of the maximum of the first 3600 pixels.
This effectively detects the positions of all of the edges. The difference between edges is
used to find the average line size. However, if the location difference must be within
50% of the initial approximate line size or it is discarded as a blank line, which will skew
the average. Figure 8.7 shows the analogous raw data profile; the blank is evident around
20,000 pixels and again around 60,000. The effectiveness of the edge detection is also
shown, since the edge detection used for the sub-sampled data set is directly used to
determine edges in the full-resolution data set.
The data is subsequently divided into average-length rows of a 2-D image.
However, because this is a sub-sampled image, the actual row length includes some
fraction of a pixel that skews the image (Figure 8.6). For example, if the actual row
length of the 25 Mpixel image is 63,442 pixels, the line size found using the sub-sampled
images will be 634. However, by row 350, that missing fraction causes the row to be
offset from the first row by more than 150 pixels (0.442 x 787). To compensate, each
line is shifted by an optimal fraction of the row number. The pointer that designates the
starting pixel of a row is shifted by some fraction of the row number rounded to an
integer (e.g., the pointer for row 300 is shifted by 133). Note that the shift is not a
circular shift within the row, but rather the pointer that determines the row’s starting pixel
in the raw 1-D line of data is shifted. The blanked edges prevent significant overlap
between rows.
212
Figure 8.6: Skewed images vs. properly aligned images. All four images come from the
same set of data. The left images illustrate the skewing that occurs due to total data grab
that is not a perfect multiple of the sub-sampling rate. The images on the right show the
alignment with the automatic correction included in the image preparation. Note the
minor imperfection of the alignment on the right does not affect the ability to further
automatically analyze the data.
To expedite the process, only the first, quarter, middle, and approximately last
rows are aligned. There are test cases where the edge blanking has not been aligned
between even and odd wavelength numbers (different scan directions). This is avoidable
if the user manually adjusts the edge position when initializing the system. However, for
the sake of automation, the quarter, middle, and ~last row numbers are forced to be odd
to ensure that they match the blank edges of the first row. If they any of those rows are
blank, the next row is used. It should also be noted that the purpose of this alignment
process is quick qualitative image quality analysis.
213
This raw aligned image is displayed so the user can see what the PicoScope
grabbed and troubleshoot any problems with the aiming or illumination that are obvious.
As with the frame grabber images, any segment that does not contain 5 lines between
blank lines is removed. The data is then divided into its 6 images (5 wavelengths and the
blank) which are then displayed and labeled so the user can determine the image quality
for each wavelength and make any necessary changes for subsequent images.
d. PicoScope Image Preparation from Complete Data Set
Higher resolution is required for quantitative data analysis, so the large data sets
must be converted into 2-D images as well. The “A” and “B” data files are opened in
Matlab. Again, they are still in the 1-D vector format. The first task is to ensure best
alignment is achieved between the primary and reference data (“A” and “B”). Ideally
each pixel of A./B would be 1 (notation indicates each element of A is divided by each
element of B, resulting in a vector the same size of A and B). Smaller and larger values
are equally problematic, so the absolute value of the log of A./B is taken (see Equation
47).
It is undesirable to consider regions of A containing vessel for the sake of the shift
optimization; the laser timing is the common factor between the primary and reference
images and should therefore be the factor by which they are aligned.
The most
consistently similar region between A and B is where the AOTF begins to pass the laser
beam (the light “turns on”). Therefore, using the sub-sampled images (during the 2-D
image formation) to determine edge locations, a binary mask is created to include only
about 15% of each fast-scan when the light is turning on (Figure 8.7). This allows for
214
image alignment without having to consider much if any of the vessel or any other
features in the images.
Figure 8.7: Profile of a segment of the reference signal. The blue profile represents the
raw profile, and the red is an overlay of the binary edge mask applied to the raw profile to
show the regions used for alignment between the reference and primary signals. This
segment is the first 60,000 pixels after any partial lines have been removed from the
beginning of the acquisition.
All of the values in the resulting masked vector are summed:
for -hrange ≤ shift ≤ hrange:
  A 
i

mask   log
47



B
i  hrange1
i  shift 



In this case, length is the number of pixels in a data set (A and B are the
sumval( shift) 
length / 2  hrange

same size), and hrange is half of the range over which B is allowed to be shifted. This
range is arbitrary, usually set somewhere between a few hundred and a thousand
215
empirically. The values of shift are integers between ±hrange to avoid exceeding the
dimensions of B. The optimum shift value corresponds to the minimum value of sumval
(vertical axes in Figure 8.8). Only half of each data set is used for alignment for the sake
of processing time.
Figure 8.8: Plots displaying shift of the reference data with respect to the primary data
vs. the result of Equation 47. The plot on the right is the vertex of the plot on the left
with an increase in resolution. The noise in the right profile is systematic, but its high
frequency makes it insignificant for this analysis process.
The primary data set is held stationary and divided by the reference data as it is
shifted from -1500 to 1500 pixels from its origin in increments of 200. All ideal shifting
thus far has fallen within this range. For each shift value, the log of the divided data is
taken and multiplied by the edge mask and summed (Equation 47). The minimum sum
value indicates the best shift amount. The process is then repeated shifting the reference
data over a 301 pixel range (-150 to 150 in increments of 20) and again over a 31 (-15 to
15) pixel range, shifting pixel by pixel about the optimum shift value determined in the
previous step (Figure 8.2). Again, the primary data is divided by the shifted reference
data; the log is taken, the edge mask is applied, and the resulting vector is summed. The
minimum sum indicates the optimum shift. The entire primary data set is then divided by
216
the optimally shifted reference data and truncated to exclude the first and last 18,815
pixels (according to the total possible shift). The shifted reference data is truncated to
match.
Both the divided and the reference data are then respectively binned by 100, such
that one binned pixel contains the sum of 100 of the original data points. These data sets
can then be handled in the same manner as the “Small” data sets. As with the frame
grabber images, any segment that does not contain 5 lines between blank lines is
removed. The reference data set is used to initially align and separate the data into 6
images by wavelength, using the blank image to identify the starting wavelength (514.5
nm).
C. Vessel and Spectrally Neutral Fundus Identification Methods
Once the original data (from the frame grabber or the PicoScope) is split into 6 2D images by wavelength, correctly oriented, and the majority of the systematic noise
from the electronics is removed (if possible), the vessel needs to be identified. Once the
vessel is identified, the vessel can be compressed (averaged) to a single profile to which a
curve can be fit. The curve can then be used to determine the OD of the center of the
vessel, even if a glint is present.
The current analysis technique actually identifies the spectrally neutral fundus in
the midst of the vessel identification process, so it will be included in this discussion.
The spectrally neutral fundus is important to the BGO technique because it provides a
reference by which the optical density of the retinal vessels is calculated (Equation 23 in
Chapter 3).
217
There are many errors with images collected by the ROx: reflections, uneven
edges (specifically between even- and odd-scan-number images), uneven illumination,
etc.
It is therefore important to be able to select a region of best image quality.
Currently, this is best done by user input. All 5 images are aligned via a row-by-row
correlation to the 514.5 nm wavelength image (“Image 1”). Once best alignment is
achieved, the images are stacked into a single “Sum Image” which is displayed to the
user (Figure 8.9). This allows the user to see any mis-alignment of edges or issue that
might affect some images but not others.
Figure 8.9: Sum Image, which is a stack of all 5 images co-aligned. The box shows the
user-defined region of interest. Note that the region has no rows void of the vessel, and
there are no reflections or other vessels in the region. It also avoids the edge regions
where the inter-wavelength alignment is not as good.
Figure 8.9 shows regions on the edges where the blank edges are not wellaligned. It should also be noted that the edges are not perfectly straightened by the
process described above in which the PicoScope data is converted from a 1-D to a 2-D
218
data set. This does not affect the quality of the vessel alignment between wavelengths,
but it can also be tweaked manually. The user is then prompted to draw a box around a
region of interest containing the vessel and its surrounding fundus. The region should
have the best quality in the image. For instance, in Figure 8.9, the user avoided the edges
with poor inter-wavelength alignment. It is this region that is used for the remainder of
the data analysis.
The new images (i.e. the selected image segments, hereafter referred to as “the
images”) are again aligned line-by-line to the first segment via correlation. This time, a
vertical correlation is included, with an allowable shift of ±2 pixels. Before any further
vessel alignment is performed, the neutral fundus should be identified to maintain the
best fundus alignment between all 5 images. This is accomplished by normalizing the
images and calculating the standard deviation for each pixel. Each image is initially
divided by its total mean. A mask is then created that only includes pixels that are greater
than the mean, since it is undesirable to include the vessel region. This masked image is
initially considered the fundus, and the original image is divided by the average of these
“fundus” values to normalize the image to the fundus.
For each pixel position, a temporary vector is created (depthvect) containing the
normalized value of that pixel position for each of the 5 wavelengths. The standard error
of the mean (SEM) for that vector is calculated and recorded for that pixel position in the
matrix "wlsem" (short for wavelength SEM; see Figure 8.10, left image). The wlsem
matrix is then used to create a mask of the regions with the lowest SEM (those less than
the half the mean of the whole wlsem matrix: see Figure 8.10, right image). These are
the most spectrally neutral pixels. Using the spectrally neutral pixels to normalize the
219
original image segments, the process is repeated. By trial and error, it was determined
that only 2 iterations of this process are needed. The mean of the spectrally neutral pixels
in each image is normalized and used as the fundus value for each of the respective
wavelengths. This provides a potential solution for avoiding capillary beds, but it does
not actually address the issue of retinal pigmentation.
Figure 8.10: Left: image of the wlsem matrix used to determine the spectral neutrality of
each pixel. The value of each pixel is the SEM of all 5 wavelengths at that pixel. Right:
image of the subsequent neutrality mask applied to the image. The only pixels with nonzero values correspond to pixels with an SEM below the defined threshold (half of the
mean value of the wlsem matrix in this case).
Initially the standard deviation was used instead of the SEM, but the dimmer
region of the image (including undesirable secondary vessels) had lower standard
deviation values than more useful regions, so it was necessary to divide by the mean of
the depthvect. Half of the mean is a relatively arbitrary threshold value for determining
which pixels are spectrally neutral. It simply has seemed to produce consistent results in
trial-and-error testing, but this is an area of the analysis that deserves more attention and
study.
Once the fundus values have been determined, the images can be aligned such
that the vessel is vertical. This way, the vessel can be collapsed (averaged) into a single
220
row with noise and anomalous artifacts averaged out, and the profile can be analyzed.
Multiple methods have been explored for vessel alignment. Three are mentioned here
because they could possibly be improved upon to produce a successful model.
The first is simply requiring the user to draw a line down the center of the vessel.
The slope of the line is calculated and each row shifted accordingly. Because all 5
images have been co-aligned, the same line is used for all wavelengths. This works well
for straight vessels; however, blood vessels often bend and curve, and even the slightest
deviation adversely affects the calculated OD spectrum. An example of this straightening
method is shown in Figure 8.11.
Figure 8.11: Vessel straightened using the line drawn by the user (top left). Note the
slight bends in the vessel at the top and bottom. The image in the top left is a stack of all
5 images before straightening; it shows the co-alignment of the images.
The second alignment method is line-by-line correlation. The vessel is initially
aligned via the line drawn by the user. This produces an initially “straightened” image
for each wavelength.
Each image is then averaged into a single profile to which each
row in the image is individually aligned. Using the average profile avoids error due to
221
anomalies in any particular line. The central 80% of each row is then aligned to this
average profile. An example of a vessel straightened this way is shown in Figure 8.12.
This method seems to produce the most consistent straightened vessels with the fewest
mis-shifted lines for images collected with the frame grabber. However, there are a few
cases in which the user-drawn line is actually better, so the analysis from the framegrabbed images includes an option for the user to choose which of these two methods
produces the best straightened vessel.
Figure 8.12: The same vessel straightened using correlation to the average profiles from
the vessel in Figure 8.11. Note the straightened top and bottom of the vessel.
The third alignment method uses the difference between pixels across each row in
an image. Initially, simple alignment of the maximum derivative in each row was used.
To improve the process, each row is smoothed by a boxcar average to reduce noise (for
straightening purposes only). The distance between pixels was also varied and the results
explored (i.e. for each pixel, instead of taking the difference of 2 adjacent pixels, the
222
difference was found between the given pixel and, for instance, a pixel 10 places ahead).
A distance of 10 pixels seems to produce the best alignment.
However, especially with the lower resolution of the frame-grabbed images, the
occurrence of anomalous lines was too frequent for the maximum differential alone to be
the straightening criteria. Instead of a differential, division is performed across the image
in a similar fashion: a given pixel is divided by a pixel several positions ahead or behind
it. Both edges are identified by running the division in both directions, dividing by pixels
ahead, then by pixels behind any given pixel. This is similar to a vessel alignment
approach previously documented by Chapman , sans use of vessel glint for centering the
alignment92. This method has also been used for vessel width measurements91, and it
should be kept in mind for similar vessel width calculations utilized on ROx data.
A binary mask is then applied, identifying the places in which the quotient is
below the threshold (as the mask is applied from left to right). A threshold is defined
empirically and used to create a binary mask of the image: the pixel value is 1 if the
quotient is below the threshold and 0 otherwise. The binary mask is then used to
determine the shift such that the last pixel in each row with a value of 1 is lined up in a
central column. The right edge of the binary mask serves to indicate the point at which
the vessel profile begins to rise again (i.e. the bottom of the vessel profile--Figure 8.13C).
It is more desirable to align the image to the vessel bottom as opposed to an edge because
the vessels are not perfect cylinders. An approximate vessel width, w, is determined via
user input, and w/2 is used for the distance between divided pixels. This way, the
division is effectively used to find the bottom of the vessel (Figure 8.13 A).
223
Figure 8.13: Process of division-straightening a vessel. This figure corresponds to the
514.5 nm images. A: the image of the vessel division before any straightening. B: binary
mask corresponding to the divided image, including only quotient values above the
defined threshold. C: binary mask shifted so that the right edge is aligned. D: image of
vessel shifted analogous to the binary mask.
In order to compensate for different optical densities, the threshold for each
wavelength is defined in terms of the ratio of the minimum to the maximum of the
respective image. The resulting binary mask can then be used for alignment (Figure 8.13
B and C). This technique is capable of producing straightened vessels that are superior to
the correlation technique, but it is less consistent in terms of anomalous line shifting,
especially in the lower-resolution or noisier images (example shown in Figure 8.14). The
“drawn-line” and “correlation” straightening techniques are used for the remainder of the
224
analysis, though the division method could be re-visited and possibly improved in the
future.
Figure 8.14: The same vessel straightened using the division method in all 5
wavelengths.
The compound straightening method requires an amount of user input that will
hopefully be minimized in the future. After selecting the initial region of interest, the
user is asked to draw a line along the center of the vessel. The user is then shown images
straightened by both the drawn line and by subsequent correlation, and the user can then
choose which method produces a more consistently straight vessel. While the correlation
method is consistently better for high-quality images with high-OD vessels, lower-quality
images may not be well suited for line-by-line correlation due to noise and/or defocus of
the image. The correlation method is improved upon by allowing the user to select a
specific region over which to correlate the lines. Additionally, the user is asked to select
a region which contains no other vessels (Figure 8.15). This step is useful for isolating
225
vessels in images that contain an artery-vein pair or forked but parallel vessels that run
close together.
Figure 8.15: Regions of interest selected by user for correlation (left) and for taking the
vertical average for the vessel profile (right). The images are different. Note the user
selects only the vessel region with very little fundus when selecting for the correlation.
The user selects the largest uniform region for the average profile.
Once the vessel is straightened, it can be averaged vertically into a single line
profile. This produces a less noisy profile than any single row, which is more conducive
to fitting a curve than using the profile from each row individually. The next step is to
identify which portions of the profiles can be considered “vessel” and which are fundus.
It is particularly important to identify the fundus for further profile normalization
described in the next section.
Again the vessel edge must be identified. Because the profile is significantly less
noisy due to averaging, the division edge identification method is much more consistent
in this case. The user is prompted to determine an approximate vessel width (w, in
pixels), which is then varied and used to determine the regions of the profile that are
vessel and fundus. Note that this approximate width is independent of the variable vessel
226
width used in the vessel fit equation—this approximate width is used primarily to
discriminate between fundus and vessel regions of the profile.
Figure 8.16: Example of edge detection method. The actual vessel profile is a solid
black line. The dotted blue and red lines show the profile division in the forward and
reverse directions, respectively. The minima of those divided lines are used to define the
sections of the profile used for fitting the vessel (green dashed line) and the fundus (cyan
dashed line). The value of w is varied in order to optimize the regions used for the fits.
Note that while it shares an initial value with the vessel width used in the vessel fit
equation, w is allowed to vary separately from the vessel diameter.
Both edges are identified by running the division in both directions, dividing from
pixels w/2 positions ahead (to the right of), then from pixels behind (to the left of) any
227
given pixel (the blue and red dotted lines, respectively, in Figure 8.16). This is again
similar to a method implemented by Chapman, except the glint is not used for
alignment91, 92. The quotient minimum serves to indicate the point w/2 pixels from which
the vessel profile begins to rise again (i.e. the bottom of the vessel profile). The vessel
edges are then defined to be w pixels ahead of the left edge and w pixels behind the right
edge.
Once the vessel edges are determined for fitting, the fundus region should be
defined. To allow for the intensity rise and fall-off of the illumination spot, the fundus is
considered to begin some distance (~w/4) from the identified “vessel edges” to avoid
artifacts that might be present as the edge of the illumination spot reaches the edge of the
vessel, such as the shadow of the vessel. It is important for calculating the optical density
that the fundus be isolated from the vessel. As detailed in the next section, the vessel fit
works even if too much of the profile is selected as the vessel; however, it is important
for the fundus to be completely isolated from the vessel in order to properly remove any
linear component (“tilt”) of the profile and normalize it.
Once the vessel is straightened and the fundus and vessel regions are
distinguished, a curve can be fit to the vessel profile. There may be more effective ways
of straightening the vessel and/or identifying the fundus; this is a part of the data analysis
that should be explored further in the future.
D. Finding the Optical Density of the Vessel
In order to determine the optical density (OD) spectrum of a vessel, the intensity
values from the center of the vessel in each wavelength should be determined20. Because
of the presence of the vessel glint, this is often impossible to determine from the raw data
228
alone. One solution is to fit a theoretical vessel profile to the actual vessel profile
excluding any region containing a glint. The theoretical intensity of the fitted curve is
then used to calculate the OD of the vessel.
a. Theoretical Vessel Profile
The first step is to establish an equation or algorithm that produces an
approximate vessel intensity profile. In the past, the vessel profile has been assumed
approximately parabolic and fit with a parabola. The resulting minimum of the parabola
was used to determine the actual intensity at the center of the vessel despite the presence
of vessel glint38. This method was initially employed in the calibration analysis for Pig 6,
but it did not produce satisfactory results.
A better model of the vessel profile is needed. This requires identifying initial
parameters that affect/define the vessel intensity in an image. For the purpose of this
algorithm, the primary physical parameter of the vessel itself is vessel diameter.
Optically, absorption and scattering in the blood should be considered. For the analysis
method used in this dissertation, these are represented by two terms: loss depth and
backscatter depth. The loss depth is how far light travels into the blood before it is
absorbed (or scattered and absorbed) for a total loss of 1/e. The backscatter depth is how
far light travels in the blood before the amount of backscattered light is 1-1/e. These
components are related to the vessel intensity profile, PV, using the following equation93:
PV  e
2L
 RV 
2
x
2
4L
L b  b  e


2
 RV 2  x 2

 for -R < x < R .
V
V
PV = 1 otherwise.
229
48
The vessel radius, RV (in pixels), is used to define where the vessel begins.
Initially, the fundus is assumed to be normalized and flat. L is 1/loss depth, and b is
1/backscatter depth (both in terms of normalized pixel depth). The horizontal axis is
defined in terms of x, which is the pixel position along the profile.
Once the theoretical vessel profile has been defined, the illumination profile must
be considered. Ideally, the spot would have a Gaussian profile with the same diameter
regardless of wavelength; however, that is not the case. The spot width varies with
wavelength.
The intensity profile of the illumination spot, PS, can be defined
mathematically:
PS 
e
 x
 
 2 RS




2
for –RS < x < RS.
2 RS
PS = 0 otherwise.
49
Again, x is the horizontal pixel position along the profile. RS is the radius of the
illumination spot. Outside of the spot diameter, the intensity is set to 0.
With the theoretical vessel intensity and illumination spot profiles defined, they
can now be convolved to simulate the light returning from the eye as the laser spot is
scanned across a single blood vessel surrounded by fundus. The convolution takes place
over a given region (“axis width”), where the vessel centered and the illumination spot is
shifted across it. An example of all three profiles is shown in Figure 8.17, where they are
shown corresponding to an actual vessel intensity profile for one wavelength. The use of
this equation has proven more rigorous and correct than the fitting the vessel profile itself
to a parabola38.
230
b. Vessel Fitting by Hand
Figure 8.17: Example of GUI that allows users fit a curve to the vessel intensity profile
by hand. An actual profile is shown in black on the right, corresponding to the line
drawn on the retinal image shown on the left. The red curve is the theoretical
illumination spot, the blue is the theoretical vessel intensity profile, and green is the
convolved theoretical profile, which best fits the actual intensity profile at 514.5 nm.
Figure 8.17 shows an example graphic user interface (GUI) by which the user can
fit a curve to actual data. The spot size, vessel diameter, penetration depth of light
(corresponding to loss depth), and back reflection (corresponding to backscatter depth)
can be varied in Equations 48 and 49. In addition, the axis width, tilt, vertical scale, and
231
horizontal shift can also be manipulated such that the fit best overlays the actual data.
These latter parameters are more a matter of centering and normalizing, though tilt is
actually compensating for an illumination angle issue. In order to use this method for
spectral analysis, the user must fit a curve to one or more vessel profiles for each
wavelength and use the fundus values and minimum of the fit curves to determine the
optical density for each wavelength. This method is time consuming and not easy to
reproduce precisely.
c. Automated Vessel Fitting
The ultimate goal is to use these equations to fit a theoretical curve to actual
vessel intensity profiles in a systematic automated fashion. While the theoretical curve
described above can be adjusted by hand to fit any given vessel profile, an automated fit
requires further consideration of the physical system in order to modify and properly
bound the curve fit. For instance, with the potential presence of a glint in the vessel, the
center might need to be ignored to avoid the glint. Additionally, if the illumination path
does not go through the center of the iris, the size of the spot on the retina can be skewed.
Furthermore, the parabolic fit to the OD spectrum should always open upward, and its
minimum should be located within a reasonable range.
These are patterns and
specifications that a human brain easily identifies and makes compensations. However,
in order for an automated algorithm to be effective, bounds and conditions must be
explicitly defined.
For the automated process, Matlab is used to perform a nonlinear regression to
find the curve of best fit from given parameters. The “nlinfit” function included in the
Matlab toolbox uses the Levenberg-Marquardt algorithm for nonlinear regression. The
232
variables used include those required for Equations 48 and 49 (spot and vessel diameters,
loss depth, backscatter depth), as well as a horizontal shift for centering the curve to the
vessel. A modified version of the command allows user input of constants in addition to
the variables used determine best fit. Such constants include horizontal axis width,
horizontal shift for centering the vessel, and minimum vessel intensity (from the raw
data), though the handling of the horizontal shift should be more closely inspected.
In several of the images, there is a linear component to the profile, most likely due
to the angle of illumination on the retina when the illumination beam does not enter the
eye along the optic axis. Instead of adding variables to the equations to compensate for
this tilt (as is done in the GUI in Figure 8.17), this linear component is isolated via fundus
identification and divided out of each individual profile (Figure 8.18).
Figure 8.18: Vessel profiles for all 5 wavelengths (distinguished by line color), before
and after division by the respective linear components. The dotted black lines indicate
the fundus regions to which lines were fit. The lines of best fit for each wavelength are
shown on the left as dashed lines. The tilt-adjusted profiles on the right are simply the
raw vessel profiles divided by the respective dashed lines.
In order to isolate the linear component, the non-vascular regions (fundus) of a
given profile are used. Once the vessel has been straightened and averaged into a single
profile, and the vessel and fundus regions have been determined, a line can be fit to the
233
fundus region. The entire profile can be divided by this line, effectively removing it from
the profile. This also serves to normalize the profile. An example of this process is
shown in Figure 8.18.
Figure 8.19: Example of a vessel intensity profile that is asymmetrical, probably due to a
variation in spot size across the image.
Another variable to consider is the change in spot size across the image, resulting
in an asymmetrical and distorted vessel profile, like the one in Figure 8.19. This is also
probably due to the angle of illumination, and it can be isolated once the linear
component of the profile is removed. This issue is handled by linearly varying the spot
234
width across the image. In Equation 49, the x-axis is centered about 0. To vary the spot
width linearly, the pixels are secondarily re-numbered such that the left-most pixel is 1
(i.e. adding half the axis width, wa, to x), and the spot radius is increased by some
variable percentage, vs, of x+wa/2.
PS 


x

 

2

R

v

x

w
/
2


S
s
a


2
e
for –RS < x < RS.
2R S  v s x  wa / 2
PS = 0 otherwise.
50
The application of Equation 50 is shown in Figure 8.20. The spot profiles all
have an area of 1, which is why the height is varied. The important aspect to notice is
how the symmetry of the profile is lost similarly to that of the example profile in Figure
8.19. Data analysis is therefore performed using the convolution of PV and PS from
Equations 48 and 50.
Figure 8.20: Effects of linearly varying illumination spot size across the vessel profile.
Left: standard spot profile (blue) compared to spot profiles whose widths vary linearly
from left to right by a factor of vs. Right: the corresponding convolved vessel profiles.
In order to use nlinfit.m command in Matlab, a function had to be created that
takes the constants, variables, and x-axis values and produces a corresponding vessel
235
profile. The variables are then allowed to change to produce a curve of best fit via the
pre-existing nlinfit.m function. However, in order to produce consistently reasonable
results, the variables must be bounded (Table 8.1). Most of these bounds have been
determined through reasoning and understanding of the system, though several were
determined empirically.
It is assumed that the user selected a region of the image that includes the entire
width of the vessel. Therefore the spot width and vessel width should both be between 0
and the width of the profile (axis width). Loss depth can never be negative, and it should
not be infinite, so the value is constrained to be greater than 0 but less than 67 (the upper
bound is arbitrary). Similarly, backscatter depth is bound between 0 and 200, with the
upper bound being similarly arbitrary. Assuming the vessel is approximately centered in
the profile, the shift that centers the fit on the vessel should be bounded by ±RV. The
linear change in spot size across the profile is bound such that the vs is between ±100
percent, but the constraints are empirical and relatively arbitrary.
Variable
Lower Bound
Upper Bound
Step Size
Default
Vessel Diameter 0
axis width
0.25
40
Loss Depth
0
67
0.05
5
Backscatter
Depth
0
200
0.05
10
Horizontal Shift -Vessel Radius
Vessel Radius
1
0
Spot Diameter
0
axis width
0.1
16
Spot Variation
-100%
100%
0.01
1
Table 8.1: Variable parameters used in the automated data analysis (in units of pixels,
except for the percentage of spot variation).
One feature of the nlinfit.m code is the ability to control the step size of each
variable during the regression. This effectively allows for resolution control. These
values were determined through working repetitively with data sets, performing fits by
236
hand. The bounds set on each variable ensure that the variables are stepped through the
regression within a reasonable range. Rather than stopping the analysis every time one of
the bounds is violated, the values are forced back to a default value that is empirically
based on common values that produce a reasonable vessel profile. The step sizes and
default values for each variable are listed in Table 8.1. Forcing the default is not as
effective as better defining the bounds, and this should be studied further in future work.
There are a few other constraints that are part of the convolution function itself.
For example, the vessel profile should not be inverted; if the theoretical profile fit is
somehow inverted, that combination of variables is ignored. Another constraint is on the
minimum value of the curve fit. If the bottom of the vessel profile is incorrect, it is most
likely due to the presence of a glint; i.e. the bottom is not low enough. The purpose of
averaging the straightened vessel into a single profile is to reduce noise in the profile, so
there should be no severe spike driving the minimum down erroneously. Therefore there
should not be a case where the curve fit should have a minimum greater than the
minimum of the actual vessel intensity profile. If a variable combination produces a
curve with too great a minimum, the loss depth variable is decreased by an empirically
effective step value of 0.8. The loss depth is chosen because its variation between
wavelengths is more significant than that of backscatter. If the loss depth can no longer
be decreased, the entire variable combination is ignored. It should be noted, however,
that in some cases the noise in the profile is still too great for this constraint to be correct.
The human eye data shown in Chapter 7, for example, sometimes has downward spikes
that erroneously drive down the vessel intensity.
237
The nonlinear regression is used to find the best curve fit to the vessel intensity
profile by optimizing the variables listed in Table 8.1.
Once the regression fit is
performed for each wavelength, the ODs for each wavelength are determined using
Equation 23. The vessel values are the minima of the 5 normalized fits, and the reference
values are the normalized “spectrally neutral” fundus values. A parabola is fit to the OD
spectrum, and the R2 value of that parabolic fit is determined and stored along with the
OD’s and minimum of the parabola (bottom plots in Figure 8.23).
d. Iterative Process of Automated Fitting
It has been repeatedly observed that the modified nlinfit.m function alone, while
useful, does not consistently produce fits of the quality required for the blue-green
oximetry technique; further specifications must be provided. These are imposed via
several iterative loops. It is probable that this methodology can be improved upon, and
future versions of analysis automation should attempt to do so.
The iterative
optimizations include glint region removal, vessel/fundus region optimization, and
freezing variables from Equation 48 that are not wavelength dependent. All of these
initially use the R2 value of the resulting parabolic fit to the OD spectrum as the metric of
best fit: the final OD spectrum has the highest R2 value. However, to further improve the
calculation, the process is then repeated in its entirety using the R2 value of the vessel
curve fits as the best-fit metric, regardless of the quality of the parabolic fit to the OD
spectrum.
Because of the possibility of a glint, this nonlinear regression process is repeated
several times, removing a varying fraction of the center of the vessel profile used for the
regression (0-50% of the width of the “vessel” segment of the profile). Each time, the
238
OD’s, minimum wavelength, and R2 value of the parabolic fit are compared. The fits that
produce the best R2 value are kept, with the stipulation that the parabolic fit opens
upward and the minimum wavelength is somewhere between 450 and 550 nm (this range
is chosen empirically until an acceptable calibration line is determined).
Figure 8.21: Fits to a vessel profile using the majority of the profile (left) vs. a smaller
segment of the profile (right). The region right of the dashed black line is used for the
vessel. The dashed cyan line indicates the fundus region; in the case of the left set of
graphs, the left and right endpoints are used to determine any tilt. Note the difference in
fits, minimum wavelengths, and the R2 values of subsequent OD spectra, even though the
vessel profiles are the same.
In working with the data, it was observed that the R2 values of the parabolic fits to
the OD spectrum could be further improved by optimizing the regions of the profile used
for vessel and fundus fits, respectively. Since the value that determines the vessel and
fundus edges is based on user input, w, it is helpful to have a means of correction and
optimization. Therefore the average profile analysis process is repeated with varying
values of w: values within ~25% of the original user input are used.
Figure 8.21
demonstrates the significance of the difference between fits when slightly different
239
segments are used. Note that this is another area where automation and optimization
could be improved by better (more precisely) defining the vessel edge and width.
Though it increases the computation time, this allowed variation improved the R 2
values such that better than 87% of the image analysis from the original calibration
experiment (Pig 6) have parabolic spectral fits of R2 > 0.9 (including 5 analysis runs of
the each image set that produced any reasonable OD spectrum), and 75% of the image
sets produced parabolic spectral fits with R2 > 0.97. The fundus region variation affects
the straightening and normalization of the vessel profile for fitting. This vessel region
variation ensures that the entire vessel is considered but allows for omission of spurious
fundus reflections, shadows, and other features; this process is designed to effectively
optimize the region of interest. In an ideal imaging situation, this would not be a
necessary consideration; however, this step has empirically been proven significant.
Note that sometimes, a particular profile segment will produce grossly spurious
data; it might therefore prove beneficial to look at the mean or mode of the minimum
wavelengths produced over the range of profile segments used. Also, the current analysis
allows for variation in the regions of the profile used as vessel and fundus; in the future,
an additional loop that allows for variation of rows included in averaging the straightened
vessel into a single profile would be potentially beneficial, as well. However, without
further optimization of code and computers, the time cost is currently too great to add it
to the analysis.
Another iteration of the curve fit is useful, this time with backscatter and RV terms
held constant. When considering the physical system, it is apparent that the actual vessel
radius, RV, should remain constant over all wavelengths. Similarly, backscattering should
240
not change very much between wavelengths (relative to the absorption), and its effect on
the curve fit is nearly indistinguishable from that of loss depth. Using this reasoning in
the algorithm requires determining the constant values of RV and backscatter. The values
of the variables are saved from the 1st-iteration curve of best fit for each wavelength. The
average vessel radius from all 5 wavelengths is calculated and used as the constant RV.
The same is done for the backscatter term. With the vessel width and backscatter held
constant, the OD spectrum, minimum, and corresponding wavelength are again
determined by the best R2 value of the parabolic fit.
It is apparent, however, that using the R2 value from the spectral fit does not
always produce the best fits to the vessel profile; in fact, some of the vessel fits have been
entirely unbelievable. A true vessel fit should produce a true OD spectrum, and a true
OD spectrum is assumed parabolic. Theoretically, both metrics should therefore yield the
same minimum wavelength. In reality, though, the parabolic shape of the OD spectrum
is not perfect, especially at extreme saturations (0 and 100% SO2). With that in mind, the
metric based on the quality of vessel fit would, in the ideal case, be a better metric than
the quality of the parabolic fit to the OD spectrum.
241
Figure 8.22: Vessel fits resulting from different metrics. Left: R2 of parabolic fit to OD
spectrum is used as metric. Right: R2 values of vessel fits are used as metric. Note the
differences in the resulting parabolic fits: 0.998 vs. 0.909. “Min. Wavelength” is the
wavelength corresponding to the minimum OD.
The R2 values of the vessel fits were then used as the metric of best fit instead of
the R2 values of the spectral fit for the entire analysis process. While the vessel fits were
largely improved, the OD spectra were not always believably parabolic. Figure 8.22 is an
example of this situation. Note that this is not even the most extreme case. Ideally, both
should produce the same results: good vessel fits and a parabolic OD spectrum (Figure
8.23). It therefore seems best to combine methods and take the average of the two
resulting minima and corresponding wavelengths.
242
Figure 8.23: Vessel fits and corresponding OD spectral fits. Left: R2 value of parabolic
fit to OD spectrum is used as metric. Right: R2 values of vessel profile curve first are
used as metric. “Min” is the wavelength (in nm) corresponding to the minimum OD.
There are several other modifications that could possibly improve upon this OD
spectral analysis in the future. The equation for the curve fit could be further explored.
Backscattering spectra could be further explored and incorporated, as could relative spot
diameter ratios measured through focus (addressing axial chromatic aberrations). The
bounds and constraints should be studied and better defined in order to improve the
accuracy and robustness of the analysis. Better definitions and/or more constraints to the
algorithm are also needed in order for the analysis to require less user input.
243
E. Determining SO2 From the Vessel Images
The ultimate purpose of measuring the OD spectrum of retinal vessels is to
determine the SrO2 (retinal oxygen saturation). The SrO2 corresponds to the SO2 of the
central vascular system.
By determining the SrO2 with reasonable accuracy, retinal
oximetry could be able to replace invasive diagnostics in many instances.
In order to calculate the SO2 from vessel images, a calibration line must be
available. The minimum value of the parabolic fit corresponds linearly to the SO2 of
blood drawn from a central line and measured with a CO-Ox20.
Once that line is
characterized by an equation, determining SO2 is simply a matter of plugging in a
minimum wavelength value and calculating the corresponding SO2.
A preliminary calibration line for the ROx is discussed in Chapter 5.
The
wavelength corresponding to the minimum OD for a retinal artery is plotted against the
corresponding SaO2 values measured via a CO-Oximeter.
The current standard for
comparison is the slope of the off-axis intravitreal illumination calibration line measured
in vivo20. Compared to plots using only the individual metrics (quality of vessel fit or
quality of OD spectrum fit), the plot of the averaged values of the wavelength with the
minimum OD has a greater R2 value, and the slope is closer to that of the intravitreal line.
Though the exact relation between SvO2 and SrvO2 is not known, the combined method
plots compare similarly to the plots from the CO-Ox data.
a.
Error Analysis
In order to quantify the reliability of the SO2 values obtained via ROx data, the
uncertainty must be considered. The SO2 calculations are based on results of best fit to
data points. One way to evaluate confidence in the OD and wavelength values derived
244
from the fits is by a Student t-test, using the sum of squares of the residuals (SSres) or the
R2 value of a fit. However since one analysis approach is based on the goodness of the
vessel fit and the other analysis approach is based on the goodness of the parabolic fit to
the OD spectrum, it is non-trivial to quantify the error of the combined process. The two
results are independent of one another; the mean of the two minimum-OD wavelengths is
currently the method of choice in terms of trying to quantify the SO2, so it stands to
reason that the error of the mean is an acceptable error metric. The final minimum OD
value as plotted against the known SaO2 for the calibration line is the mean of minimum
values found using two different analysis methods that should ideally yield the same
information.
It should be noted, however, that with improved image quality (i.e. better glint
removal and vessel fitting in addition to the resolution improvements and more precise
control of a calibration experiment used in Pig 7), only one of these methods should be
needed, and the error propagation will be simpler. Once a calibration line is firmly
established, the R2 value of that linear fit could be used to quantify error and/or
confidence in a measurement. In previous work, the standard deviation of the calibration
residual error was used to quantify uncertainty in the %SO2 measurement20.
b. Error Analysis for Future Consideration
In the past, SEM of the vessel and fundus values for each wavelength has been
used to determine the error of the OD measurements. Using the slope of the off-axis
intravitreal illumination calibration line, the SO2 error was then determined as 3.03 times
the error for the wavelength corresponding to the minimum OD20,
47
.
Now, when
considering the best-vessel-fit method, the SEM of the fundus values can still be used,
245
but the R2 values of the vessel fits might be used to quantify the error in the vessel
measurements with the Student t-test. However, for the best-spectral-fit method, the R2
value of the final parabolic fit might be a better means of quantifying confidence in the
measurement, since it is the metric through which the minimum OD value is determined.
In fact, for either method, it could be useful to quantify the goodness of fit of both vessel
and spectral fits.
The R2 value is the coefficient of determination and can also be defined as the
squared correlation coefficient. For the sake of error analysis, the standard error of the
correlation coefficient (R, where R2 is the coefficient of determination) is defined as
1  R2
SER 
n2 ,
51
where n is the number of observations94. The number of data points used to fit the
vessel profiles are a couple of orders of magnitude more than the 5 points used for the
OD fits, so these vessel fits contribute a relatively insignificant amount of error using this
method.
Since SEOD is ~3 times the error of the minimum wavelength, and the limit for
allowable error is considered to be ±3% SO2, the allowable error in the minimum
wavelength value is ±1 nm. For the Student t-test, this translates to Δx = 1. As an
example, let the OD spectral fit have an R2 value of 0.9 for a fit to the 5 OD data points.
This is then directly plugged into the t-test equation,
x 
SE  t
n 1 .
52
Using SER as the standard error in this example and plugging values and Equation
51 into Equation 52, the resulting value for t is 10.95. Using a look-up table, this equates
to approximately 99.8% confidence that the wavelength corresponding to the minimum
246
OD is within 1 nm of the actual value. Another way to think about the error is in terms of
a minimum confidence value. For example, in order to have 99.9% confidence that the
error is within 1 nm, t must be at least 12.924. This can be plugged into the following
inequality:
12
53
t2 .
It follows that R2 must be better than 0.928 to meet this requirement. For the
R2  1 
vessel fits, the number of points used is so large that the errors of the R 2 values are
several orders of magnitude lower than the error of the OD spectral fits.
F. Conclusions
An automated data analysis procedure is crucial for expedient data analysis. The
algorithms should be able to handle images containing vessels of varying quality and
orientation. The results should be consistent, accurate, and reasonable. Required user
input should be minimized. The data analysis procedures described in this chapter show
improvements to these ends.
Most analysis methods used for retinal oximetry use average values, often over
many pixels, to determine OD values20, 35, 47, 90, 91, 95. Some groups simply omit data that
contains glint51, 90, but they are using fundus camera imaging that collects large amounts
of vessel data over relatively long periods of time. The nature of the data collected by the
ROx does not allow for such omissions, though physical removal of the glint would
improve the quality of analysis significantly (discussed further in Chapter 10). The
analysis methods used on the ROx are similar to those used by Smith et al. 38, 48, using the
vessel profile to determine the vessel OD. A novel theoretical set of vessel profile
equations are introduced and utilized (Equations 48 and 50) on ROx data gathered from
247
pig arteries (Chapter 5), pig veins during sepsis (Chapter 6), human retinal vessels
(Chapter 7), and retinal vessels in enucleated pig eyes, with varying degrees of success.
Vessels can be targeted and imaged with the ROx, and 2-D images can be
collected and co-aligned. The spectrally neutral fundus region can be determined with
minimal user input, though the spectrum of the fundus (especially in/between humans)
and identification of spectrally neutral regions deserves further study.
A user can
consistently and relatively simply isolate and straighten a vessel, from whence a single 2D vessel profile is produced for each wavelength. A curve can be fit to the vessel profile
automatically via several iterations of nonlinear regression, though there is significant
room for improvement when dealing with images containing significant vessel glint.
A relatively consistent method of determining the spectral minimum has been
identified, though there is much room for improvement. With the values of the spectrally
neutral fundus and the fit to the vessel profile, the OD spectrum can be automatically
generated. Using both the quality of the vessel fits and the quality of the OD spectral fits
as metrics seems to produce a reasonable minimum OD value whose wavelength could
possibly be translated in terms of SO2 with further calibration. While data collected thus
far has not been of a quality conducive to rigorous testing of this method, the analyses
performed on data gathered up to this point consistently support this theory that
combining the two methods is superior to either of them used alone.
Future work should include revisiting the error analysis methods, as well as
improving the expediency and user-friendliness of the code. With improved image
quality, the OD analysis and vessel fit methods should be re-considered. The model
vessel profile equation should also undergo further analysis and improvement, though it
248
is a significant improvement upon the parabolic fits used in the past38. The contributions
of the vessel diameter might be considered further63, 90. The contributions of scattering
spectrum, as discussed in Chapters 5 and 10, should possibly be better incorporated into
the automated analysis as that interaction is more clearly understood.
249
Chapter 9:
MINIATURIZATION OF THE ROX
A. Introduction
In order to make the ROx a viable instrument for use in medical facilities, it must
be made small enough to be transported, while at the same time maintaining or improving
upon the functionality of the current device. Any optical system can be made smaller,
but at a cost in more expensive components and tighter tolerances. This chapter contains
an optical design study to explore concepts for the miniaturization of the ROx optics.
Section B explains the requirements, Section C a description of a miniaturized optical
design, and Section D the consolidation and miniaturization of the rest of the ROx
system. These changes are not currently realized; they have been investigated to help
guide further development of the ROx.
B. Requirements for the ROx as a Medical Device
Chapter 4 outlined the requirements of the ROx. At least 5 wavelengths in the
blue-green spectral range are required for BGO, in addition to an IR wavelength for
targeting. Two-axis scanning for imaging requires two perpendicularly scanning mirrors:
a fast scan mirror (~4 kHz) and a slow scan mirror (~10 Hz). Both mirrors must be
conjugate with each other as well as the pupil of the patient’s eye. The beam itself should
be very nearly collimated as it enters the eye, with adjustable power to optically
compensate for patients with less than perfect vision—at least for myopia and hyperopia.
Note that unless otherwise stated, the eye is modeled as a single perfect lens with an EFL
250
of 16.78 mm. This way, any aberration from the eye model itself is propagated into the
system analysis.
The return path should keep the field stop filter (currently the field stop wire
filter) confocal with the retina. The light flux in the return path is the priority and should
be maximized, even at the expense of power lost in the illumination path. Another
consideration is the amount of scattered light from the retina that reaches the detector.
A practical limitation is the 20° angular range of the fast scan mirror, with the
minimum scanned region of at least 500 x 500 μm. A useful merit function for this
device is the geometric spot size; it ultimately determines physical limit of the spatial
resolution of the images generated by the ROx. However, the shape of the spot and
amount of aberrations present are important as well.
C. Miniaturized Optical Design
Development of an optical design that meets requirements and maintains similar
optical qualities as the pre-existing design starts from and understanding of the chief and
marginal rays. The Lagrange invariant, Ж, is related via the chief and marginal rays by:
  nu y  nuy
54
.
This relationship holds for the paraxial rays at any point along the optical axis.
The chief ray is represented by the paraxial angle (ū) and height (ȳ). Similarly, the
marginal ray is represented by u and y96. The ROx system is in air, so n is 1. The chief
and marginal ray heights in the orthogonal transverse direction are denoted by x and x,
respectively.
The miniaturized design should behave similarly to the current system. In the
case of the ROx, it is desirable to design a short optical system with an equivalent Ж.
251
The “y-y bar diagram”97 is useful for paraxial design. For a given optical system, a ray
trace is performed. A plot is generated of the ray heights at each optical element in the
system, sequentially, plotting chief ray height vs. marginal ray height. The y-y bar
diagram for the current configuration of the ROx (Chapter 4) is shown in Figure 9.1.
Figure 9.1: Ray trace plot (y vs. ȳ and x vs. x ) for the current ROx configuration from
pellicle to eye (forward illumination path).
In the figure, the shaded region represents the area between OAP 1 and OAP 2
swept by a line from the origin. The point labeled “Iris” is the power of the eye
represented by a single perfect lens of EFL 16.78 mm. The ray trace was performed in
Code V with an approximate beam size of 1 mm (entrance pupil diameter) without
scanning. Commands “RSI 0 1 0 0” and “RSI 0 0 0 1” for chief and marginal ray
heights, respectively, in the y-direction; similarly “RSI 1 0 0 0” and “RSI 0 0 1 0” were
252
used for the x-direction. For the chief rays, fields with arbitrarily chosen object angles of
0.57 were used. The stop used for his plot is the FSM to show the conjugate nature of the
SSM and FSM, specifically. The working distance is 101.6 mm, and the Lagrange
invariant is 0.0149.
An important characteristic of the y-y bar diagram is that area between elements
swept by a line from the origin is proportional to distance between elements along the
axis (e.g. the shaded region of Figure 9.1), thus reducing the area inside the polygon will
reduce the size of the system. The y-y bar diagram indicates the optical elements with
largest contributions to the length: the position of the pellicle, and the powers of OAP 1,
OAP 2, Sph 3, OAP 4, and L 3 (OAPs 1 and 2 being the greatest contributors). Figure
9.1 shows the return case, wherein the iris of the eye is the aperture stop; however,
because the location and size of the iris are not fixed within the system, the aperture stop
of the return path within the ROx itself should be determined. Though it is not shown
here, the aperture stop for the illumination spot is a pinhole before the pellicle (refer to
Chapter 4), which is not even a component of the return path. Instead, the return path of
the scattered light from the on-axis beam is shown.
The plot shows which elements are the aperture stop and its conjugates, because
these are located on the chief ray (horizontal) axis; the stops and any conjugates occur
where the chief ray crosses the optical axis. Also, vertical and horizontal lines on this
type of plot indicate collimated beams and telecentric beams, respectively.
In this
system, the x- and y-axes behaves in a similar fashion, which is also shown in Figure 9.1.
The fast- and slow-scan mirrors (FSM and SSM) and the iris of the eye have chief
ray heights of nearly zero, since they are located at pupils. In order to determine which
253
surface is the aperture stop for the scattered return beam within the ROx, the semiaperture (a) of each element, k, is divided by its respective marginal ray height. The
element with the minimum value is the aperture stop. Note that the marginal ray here can
be any trial ray starting from the axial point on the object. Only the y-direction is shown
here, since the x- and y-directions are comparable.
astop 
ak
yk
55
min
.
For the current system, the data for this calculation is shown in Table 9.1. Figure
9.1 already takes these results into account, showing the FSM as the aperture stop. Note
that the data used in Table 9.1 is taken from the on-axis illumination beam. The scattered
light in the return path will over-fill it as one would expect at an aperture stop.
Surface Semi Aperture min|a/y| = Stop Surface (On-axis beam)
in mm
|a/y|
APERTURE STOP for return beam within ROx
FSM
4.625
9.25
system itself
OAP 1 25.400
50.68
OAP 2 25.400
51.30
SSM
12.700
25.65
Sph 3
12.700
25.68
OAP 4 25.400
26.63
L3
12.700
14.08
L4
12.700
12.06
Iris
4.000
3.83
APERTURE STOP
Table 9.1: On-axis beam path data for locating limiting apertures and stops. The iris is
the aperture stop for the entire imaging system. It is also the exit pupil, and the fast-scan
mirror is the entrance pupil.
Thus far only the beam path of an instantaneous, on-axis spot on the retina has
been considered. The beam size of any instantaneous spot during scanning should not be
significantly affected while scanning (for the sake of locating stops and apertures), so the
254
scanning has been ignored in this case. However the behavior of the entire scanning
illumination path must also be considered. The nature of this design is such that the
marginal ray of the scanning path is geometrically similar to the chief ray of the
instantaneous spot on the retina, and vice versa (e.g. the scanning path has focused nodes
where the instantaneous beam is collimated; see Chapter 4, Figure 4.1). The aperture
stop of the scanning path is found using the same technique used for the on-axis spot
path, but using its chief ray instead. The results are shown in Table 9.2. The first offaxis parabolic mirror (OAP 1) is the aperture stop (with OAP 2 effectively imposing the
same limits). This means the aperture size of the first two OAPs is important to consider
in the design—they must be large enough to handle the desired angular range of the
scanning mirrors.
Surface Semi Aperture
min|a/ȳ| = Stop Surface (Scanning beam)
in mm
|a/ȳ|
FSM
4.625
1020.97
OAP 1
25.400
17.63
APERTURE STOP for scanning
OAP 2
25.400
17.65
SSM
12.700
3377.66
Sph 3
12.700
28.23
OAP 4
25.400
31.22
L3
12.700
106.63
L4
12.700
65.91
Iris
4.000
Infinite
Table 9.2: Scanning illumination path data for finding stops. The chief ray of the
illumination beam path (no scanning) is simply a scaled version of the marginal ray of the
scanning beam path, so it is used here. For the current configuration, the first parabolic
mirror is the aperture stop.
A person with perfect vision should be able to relax the eye, focusing at a distance
(infinity, effectively), and the beam will be focused to the retina. Moving L3 focuses the
optics to compensate for myopia or hyperopia.
255
The overall length of the system, represented by the area inside the y-y bar plot,
can be reduced by shortening the focal lengths of OAPs 1, 2, and 4. The focal lengths of
L3 and L4 can be shortened as well. It is important, though, to maintain the system pupil
locations and a similar beam shape.
The choice of optics currently used in the ROx has many ramifications on the
design. To avoid ghosting and chromatic aberrations associated with lenses, this design
primarily uses mirrors. The uniformity and size of the spot illuminating the retina is a
concern, so OAPs are utilized in the current design to image the stops onto one another
because they introduce less aberration than spherical components. Due to the bends in
the beam path, the spherical symmetry would not be preserved; spheres would introduce
coma and astigmatism into the system. Paraboloids are used for converting between
plane waves and spherical waves while minimizing coma and astigmatism, which is
precisely what this system requires. 90° OAPs are common off-the-shelf optics, which
are not expensive, so they are used here. The optics are also easier to align when bends
occur at right angles (i.e. the optical axis is bent 90°; this is referred to as θOA in Figure
9.2). However, because all of the axis bending occurs in the direction of the slow scan
(arbitrarily the y-direction in this design), some astigmatism is introduced. This was
empirically remedied in the current design by replacing OAP 3 with an on-hand spherical
mirror of the same focal length.
Toroidal mirrors have two different radii of curvature in orthogonal directions and
are relatively easy to fabricate. Consider the case where the optical axis is perpendicular
to the center of an OAP. If the diameter of the mirror is small compared to its radius of
curvature, it is well approximated as a toroidal mirror98. In light of this simplification,
256
toroidal surfaces closely approximate parabolic surfaces over limited regions, making
them a possible choice for controlling astigmatism using an aspheric surface. For an onaxis beam, OAPs are excellent for compensating for astigmatism. However, the scanning
mirrors create the off-axis beam, resulting in astigmatism that is a function of field of
view (i.e. scan range). Toroids can be designed to produce compensating astigmatism for
the off-axis beam.
Figure 9.2: Diagram of an off-axis paraboloid (OAP) with respect to the paraboloid on
which it is based (the “parent” paraboloid). This shows the distinction between the focal
length of the paraboloid (f) and the back focal distance (BFD). The blue lines represent
the incident and reflected optical axes. The red lines are rays traced from infinity. The
focal lengths of the complete “parent” paraboloid and of the OAP are shown, as well as
the angle θ used in Equations 35 and 57 and θOA used in defining the OAP. Note that,
corresponding to the convention of these designs, this diagram is in the y-z plane, where
the optical path propagates in the z-direction.
257
An OAP has two definitions of the focal length to consider: the “parent” focal
length (f) refers to the focal length of the vertex radius of curvature of the paraboloid of
which the OAP is a segment (f = R/2, where R is the vertex radius of curvature).
Additionally, the “reflected” focal length (the back focal length, BFD) refers to the
distance from the center of the mirror to the focal point. Figure 9.2 illustrates the OAP
mirror segment with respect to the base paraboloid. For a 90° OAP, BFD = 2f and θ =
45°. When designing a toroidal mirror approximation of an OAP, the tangential and
sagittal radii of curvature depend on the reflected focal length, f, of the paraboloid and the
angle, θ, between optical axis and the normal line extending from the center of the OAP
(see Figure 9.2)99. In this design, cy is the tangential curvature since the bending of the
optical path occurs in the y-direction:
cy 
cos3  cos3 

for 90° OAPs.
2f
BFD
56
The sagittal curvature is given by cx:
cx 
cos cos

for 90° OAPs.
2f
BFD
57
The respective radii of curvature are simply the reciprocals or the curvatures.
It is important to consider the size of the optics and how they will be mounted in
relationship to one another. Many off-the-shelf OAP’s are fairly large, extending towards
the bottom of the parabola. The scanning mirrors must also be considered, as they extend
in the direction perpendicular to their respective scanning. The angle of the pellicle must
also be taken into account.
258
a.
Design Process: Illumination Path
With all of these considerations in mind, a miniaturized design was developed in
Code V. With the current system as a template, the object is at infinity and the entrance
pupil diameter is 1 mm. During the design process, the system stop was placed at the
pinhole position 50 mm in front of the first optical element because the laser line itself
must be cleaned up. As with the previous design, the surface on which the optics are
mounted is defined as the y-z plane.
Figure 9.1 indicates how the system size can be reduced by shortening the focal
lengths of OAP 1, OAP 2, Sph 3, and OAP 4. Given the sizes and focal lengths available
off the shelf as found in an online search, and considering space needed for mounting, the
focal lengths of the first two focusing mirrors were reduced to approximately a quarter of
their original lengths and the second two were reduced to half of their original length.
All mirrors are also positioned to bend the light path at right angles. This is both because
of the use of 90° OAPs and for simplifying alignment.
The first three focusing mirrors have a 25 mm BFD, and OAP 4 has a 50 mm
BFD. The focal lengths of the respective paraboloids are 12.5 mm and 25 mm. In Code
V, the vertex radius of curvature is used. According to the thin lens equation, the radius
of a mirror (R) with respect to its focal length is R = 2f (where index n = -1 for reflection
in air). Therefore the respective vertex radii entered into Code V are 25 mm and 50 mm.
Because these are parabolic components, they have a conic constant of -1.
The astigmatism introduced by the bending of the optical axis is accounted for in
the miniaturized design by replacing OAPs 2 and 3 with toroidal mirrors. These are also
used to compensate for the asymmetry introduced by the scanning mirrors.
259
Using
Equations 35 and 57, the radius of the y-curvature for both toroids should be 70.71 mm,
and the radius of the x-curvature should be 35.36 mm to approximate the OAPs. To
simulate scanning, the system was set up with 5 “zooms”, or positions, with the scan
mirrors at different angles. Zoom 1 simulates the minimum scan angle for both the SSM
and FSM, Zoom 3 simulates the on-axis beam (as if there is no scanning), and Zoom 5
simulates the maximum scan angle for both the SSM and FSM. Zoom 3 is used as the
reference to characterize the system.
The distance from the pellicle to the FSM was also shortened while maintaining
the optimal 26° reflection of the beam into the system as discussed in Chapter 4.
However, in order to allow room for OAP 1 and the FSM mount, which extends in the ydirection, the pellicle is angled in the x-direction, allowing the light to be brought into the
system perpendicular to the progression of other optical elements (see Figure 9.7). An
alternative solution would be to switch scanning axes so the first scan occurred in the ydirection and the second one occurred in the x-direction, since the scan mirror mount
extends perpendicularly to the scan direction. While feasible, this solution was not
explored further. The light path is bent in the y-direction, so y-scan is introduced as late
as possible in order to minimize aberrations that are dependent on field height, like coma
and astigmatism.
260
Figure 9.3: Y y-bar diagram showing the optical elements plotted sequentially in the
forward direction. The chief and marginal ray heights in both the x- and y-planes are
plotted. The FSM was used as the stop for this plot. The working distance (between
OAP 4 and the eye) is 53.0 mm, with the Lagrange invariant equal to 0.0149 (this is
equivalent to that of the current system).
The y y-bar diagram for this miniaturized design without focusing lenses is shown
in Figure 9.3. Because two of the OAPs are replaced by toroidal mirrors, the x-axis is
also plotted to verify that the beam behaves similarly along both scan axes and to
compare it to the current configuration. Note the collimation of the system, both in the
chief rays and the marginal rays, as indicated by horizontal and vertical connecting lines,
respectively. Also notice the chief ray heights of the FSM, SSM, and iris: they are all
261
conjugate pupils, as required to maximize the amount of light scanned into the eye. It is
where the light first interacts with the eye. Using first order analysis by hand, the
Lagrange invariant is found to be 0.0149 at the OAPs 3 and 4 using Equation 54. This is
equivalent to the Lagrange invariant of the current system. Note, however, that in
addition halving all of the focal lengths, this design takes a step further and halves the
focal lengths of the first mirror pair again since they only serve as a relay with
magnification of -1.
A major function of Code V is its automatic optimization function. Several
system parameters can be set to vary. The program then optimizes them according to a
set of user-defined merit functions to maximize image quality and meet constraints like
marginal ray height or angle at a given surface, or to minimize specific aberrations. The
summary of constraints used for this design is listed in the next paragraph. The design
initially used the iris as the stop surface, since it is the aperture stop of the overall return
path. The toroidal mirrors were allowed to vary in the y-direction first. The distances
between OAP 1 and Tor 2 and between Tor 3 and OAP 4 were also allowed to vary.
The following lists the constraints used for optimization. In all cases, the 488 nm
wavelength is used with the on-axis field (object angled at 0°).
Throughout the
optimization process, close attention was paid to the spot size over all fields, as well as
the sensibility of the spacing and curvature of the components. Elements were usually
varied one at a time to hone in on the solution in a controlled fashion. The elements were
then varied in groups for further optimization.
1. Local Y Surface Coordinate of the Chief Ray targeted to equal 0 mm at the FSM
and SSM for Zooms 1, 3, and 5.
262
2. Local Y Surface Coordinate of the +Y Meridional Ray targeted to equal -0.5 mm
at Tor 2, the SSM, and Tor 3 for Zoom 3 (no scan).
3. Local Y Surface Coordinate of the +Y Meridional Ray targeted to equal 1.0 mm
at OAP 4 for Zoom 3.
4. Local Angle of Incidence of the +Y Meridional Ray targeted to equal 0° as it
enters the eye (at the surface of the anterior cornea).
5. The x-radii of the toroidal mirrors were then allowed to vary while all other
parameters were frozen. The same constraints were held, except in the x-direction
with the +X Sagittal Ray.
The iris of the eye is not an actual component of the ROx, so the system stop was
moved to the FSM; it is the stop for the ROx in the current design. The x- and y-radii of
Tor 3 were allowed to vary, along with the distances from SSM to Tor 3 and Tor 3 to
OAP 4, with the constraint of a chief ray height of 0 at the cornea in both x- and ydirections. Once ȳ and x were less than 0.01 mm at the SSM, those parameters were
frozen and the distance from OAP 4 to the eye was varied until the chief ray height in
both the x- and y-directions were less than 0.01 at the anterior cornea and iris.
In light of cost, all radii of curvature were rounded to the nearest 0.01 mm and
the elements were optimally re-spaced, which resulted in no significant change in spot
size, aberration, or ray collimation. While rounding to the nearest 0.1 mm actually
slightly improved the spot size and chief ray heights at the stop conjugates, the ±45°
astigmatism and defocus due to field curvature worsened by about half a wave in the
maximum scanning angles, which is significant.
263
To examine the design of the scanning system alone, the system is specified in
Code V as afocal (“paraxial conjugates corrected”).
To observe the spot size and
aberrations for an ideal eye, a perfect lens at the eye position is used to model the eye.
The perfect eye was specified to have an EFL of 16.78 (following the Arizona eye model
2
). The resulting spot size is shown in Figure 9.4, looking at the spot for scan angles on
axis and at Zooms 2 and 4 (±250 μm in both x- and y-directions). Figure 9.5 shows the
wave front aberration plots for both the x- and y-ray fans in Zooms 2-4. The aberrations
are greater in the y-direction, i.e. the direction in which the most bending of the optical
axis occurs. Table 9.3 shows the aberrations in terms of Zernike polynomials, following
the definitions of Born and Wolf in Principles of Optics, 6th ed. (Pergamon Press, New
York, 1989)100. The shape of the spots (100% size 13 μm to 37 μm) reflects the greatest
contributor to aberrations—the ±45° astigmatism coefficient. By comparison, the current
system (excluding focusing lenses) has a theoretical spot size (100%) ranging from 29 to
36 μm.
During optimization it was found that a better spot size can be found with this
minimized system if the magnification is allowed to be less than 2. However, in order to
gather as much scattered light as possible, the magnification of the beam in the forward
path is held at 2 (giving a magnification of 0.5 in the return path).
264
Figure 9.4: The spot size and shape for all 5 wavelengths at the corners of a 500 x 500
μm image (Zooms 2 and 4) and at the center of the image (Zoom 3) with no focusing
lenses present. Units are in mm.
ZERNIKE POLYNOMIAL COEFFICIENTS--No Occular
Number Order Value (waves at 488.0 nm)
Aberration Type
Zoom 2 Zoom 3 Zoom 4
1
0
0.1981 -0.0877 -0.1956
Piston (constant)
2
1
0.0045 0.0000 0.0011
Distortion - Tilt (x-axis)
3
0.3122 0.3314 0.2668
Distortion - Tilt (y-axis)
4
2
-0.1397 0.1626 0.2531 Astigmatism (axis at 0° or 90°)
5
0.2038 -0.0974 -0.1914
Defocus - Field Curvature
6
0.6643 0.0000 -0.7008
Astigmatism (axis at ±45°)
7
3
-0.0001 0.0000 0.0017
Trefoil (x-axis)
8
0.0022 0.0000 0.0006
Coma (x-axis)
9
0.1519 0.1671 0.1339
Trefoil (y-axis)
10
0.0052 -0.0126 0.0018
Coma (y-axis)
Table 9.3: Zernike polynomial coefficients of maximum/minimum scan angles for a 500
x 500 μm image (Zooms 2 and 4) and of the beam on axis with no focusing lenses
present. Notice that there is less than a wave of any aberration.
265
Figure 9.5: Wavefront aberration plot: optical path difference in waves for all 5
wavelengths for the system without focusing lenses.
266
b. Design Process: Focusing Lenses
An ophthalmic instrument needs to be able to focus to properly couple into the
patient’s eye. To do this, two lenses (L3 and L4) are placed at the end of the illumination
path in order to adjust for near- and far-sightedness. These lenses are not included in the
previous figures or tables. They are ideally arranged as a 4f system, so they have a
magnification of -1, maintain spot collimation, and they image the FSM and SSM onto
the iris. However, one lens (not the one nearest the patient to keep the patient stationary)
can be moved back and forth to counteract myopic and hyperopic cases, respectively. In
order to ensure the ability to rotate about the eye by ±45° in the y-z plane (for vessel
targeting purposes), the length from OAP 4 to the eye should be 75-100 mm, depending
on the shape of the optical housing.
Another consideration is to adjust the allowable axial variation of eye position; it
is necessary to accommodate a range of eye positions. Consider the case in which the
scanning mirrors are off. The perfect eye can afford to be (axially) positioned over a
large range—theoretically infinite—and still produce an ideal beam size on the retina.
However, if the power of the eye is off by ±5 diopters, that allowable range is
significantly less; the focusing lenses (f = 25 mm) used in the first year of pig and human
data acquisition had a range of about 45 mm over which the retinal spot size was less
than 50 μm. However, it only allowed 2.8 mm of error from the ideal eye position in one
direction. Those lenses were replaced with two lenses with focal lengths of 60 mm, and
the range improved to about 70 mm, and it allows 4.1 mm of error in the more limited
direction from the ideal eye position.
267
In the miniaturized system, the lenses need to be fast enough to keep the length
minimal but slow enough to allow a reasonable axial range of motion for the eye of the
patient. Instead of limiting the focal lengths of the lenses, a pair of 90° folding mirrors
has been added to the focusing system to compensate for a longer path length (hereafter
listed as FM 1 and FM 2) without introducing significant aberrations. One is located
shortly after OAP 4 and the other about half-way between the focusing lenses in order to
give L3 room for focusing motion (see Figure 9.6). This allows for using focal lengths of
50 mm for L3 and L4 (considering the current design). From FM 2 to the eye is 99.4
mm, which gives necessary range of motion about the eye for targeting. The axial
“wiggle room” for the eye is 60-90 mm for a focusing power of 50 diopters, with the
most limiting distance from the ideal position now at 26 mm.
Figure 9.6: Layout of the miniaturized ROx in the Y-Z plane. The ray coloring is such
that blue corresponds to on-axis, and red and green correspond to scan angles that give
ray heights of ±250 μm at the retina. The FSM and OAPs are drawn to scale to show the
268
space they take up in the system, illustrating the need for the beam to come in from the
X-Z plane.
A concern with the use of lenses is chromatic aberrations. The overall spot size
and uniformity were significantly improved with the use of achromatic lenses. Bi-convex
lenses were chosen to reduce spherical aberration. This design implements a pair of
inexpensive bi-convex achromats. They should have an AR coating at least in the bluegreen range to prevent ghost reflections with the laser wavelengths. The lenses are
examples of off-the-shelf lenses from the Thorlabs website.
Figure 9.7: 3D view of the forward path of the system. The optical elements are shown
as frames, and the OAPs are represented in their full parabolic shape. The ray coloring is
such that blue corresponds to on-axis, and red and green correspond to scan angles that
give ray heights of ±250 μm at the retina.
Figure 9.7 and Figure 9.6 show schematics of the design layout with ocular
lenses. The three zooms for the three design fields of view are used in these figures,
where Zooms 2 and 4 result in a beam position of (±250 μm, ±250 μm) on the retina.
Figure 9.7 shows the 3D beam path, with the optics displayed as wire outlines. Note the
position of the beam entrance point. Figure 9.6 displays not only the beam path, but the
amount of space the larger elements occupy.
269
An alternative to focusing lenses is the use of focusing mirrors, where the length
from the last optic to the eye is stationary, but one focusing mirror and a corresponding
flat folding mirror are allowed to move, effectively the same way Oc 1 is allowed to
move. This was explored in Code V for the current system using OAPs, but the bends
introduced other aberrations such that the spot quality was not actually improved overall.
However, this is a solution worth pursuing in the future, perhaps with toroidal mirrors.
Figure 9.8: Y y-bar diagram showing the optical elements plotted sequentially in the
forward direction. The chief and marginal ray heights in both the x- and y-planes are
plotted. The FSM was used as the stop for this plot. The working distance (between L 4
and the eye) is 47.2 mm.
The y y-bar diagram is shown in Figure 9.8. Notice that all of the stops are still
present.
The height between optical elements has decreased, though width has not
because the magnification of the system is the same. The resulting spot size (again using
the afocal system model in Code V) is shown in Figure 9.9. Note its similarity to the
270
spots in Figure 9.4. The focusing lenses have little effect on the spot size or overall
quality. For comparison between the system with and without the focusing lenses, Table
9.4 lists the Zernikes, and the OPD plots are shown in Figure 9.10.
In Table 9.5, the lens data is shown for the scanning system with focusing ocular
lenses in place. The Semi-Apertures of the optics were automatically chosen by Code V
in order to determine the minimum diameter required for each optic. Note that the semiapertures of the OAPs are very large compared to the other optics; this is because Code V
accounts for the entire parabola as opposed to the off-axis segment that will actually be in
place in the system. Figure 9.7 shows this as well.
.
Figure 9.9: The spot size and shape for all 5 wavelengths at the corners of a 500 x 500
μm image (Zooms 2 and 4) and at the center of the image (Zoom 3) with focusing lenses
present. Units are in mm.
271
Figure 9.10: Wavefront aberration plot: optical path difference in waves for all 5
wavelengths for the system with focusing lenses
272
ZERNIKE POLYNOMIAL COEFFICIENTS--With Occular
Number Order Value (waves at 488.0 nm)
Aberration Type
Zoom 2 Zoom 3 Zoom 4
1
0
0.2117 -0.0154 -0.1813
Piston (constant)
2
1
-0.0036 0.0000 0.0079
Distortion - Tilt (x-axis)
3
0.2945 0.2771 0.2735
Distortion - Tilt (y-axis)
4
2
-0.1379 0.1032 0.2500 Astigmatism (axis at 0° or 90°)
5
0.2165 -0.0114 -0.1777
Defocus - Field Curvature
6
0.6409 0.0000 -0.7164
Astigmatism (axis at ±45°)
7
3
0.0001 0.0000 0.0026
Trefoil (x-axis)
8
-0.0019 0.0000 0.0040
Coma (x-axis)
9
0.1479 0.1390 0.1372
Trefoil (y-axis)
0.0055 0.0053 0.0021
10
Coma (y-axis)
Table 9.4: Zernike polynomial coefficients of maximum/minimum scan angles for a 500
x 500 μm image (Zooms 2 and 4) and of the beam on axis with no focusing lenses
present. Notice that aberrations are comparable with these lenses in place; ±45°
astigmatism is the largest at about half a wave.
Table 9.5: The Code V lens data for the miniaturized system with the focusing lenses in
place and an ideal lens used in place of the eye (with about the same focal length in air).
This is an afocal system. The stop is set to the entrance point pinhole.
273
c. Design Process: Return Path
The beam returning from the eye shares a path with the illumination beam until it
reaches the pellicle. The beam traveling from the FSM through the pellicle must be
focused onto the field stop wire filter (FSWF), which passes scattered light from the eye
to the PMT. When miniaturizing the design, the size of the FSM must be considered in
relation to the incoming illumination angle at the pellicle. The FSM is a “resonance
scanner”, using a galvo motor to scan back and forth at a single given frequency.
Currently, the common resonance scanners on the market have a total scan range of 20°
at 4 kHz with a mirror diameter of 12.7 mm (Cambridge Technology). With these
dimensions, the pellicle can be placed 25 mm from the FSM (see Figure 9.6). This
allows for a shorter focal length in the optical element focusing the light from the FSM
onto the FSWF, and therefore a shorter distance from the focusing optic to the FSWF.
The dimensions of the base of the FSM (again shown in Figure 9.6) also dictate the
position of the FSWF and the PMT.
The magnification of the retina onto the FSWF is 2.8 in the slow-scan direction
and 3.0 in the fast-scan direction. In order to pass an area 900-1,000 μm in diameter
through the filter, the pinhole need a diameter of 2.8-3.0 mm. Similarly, if ~200 μm of
the retina should be blocked because of potential central glint, the bisecting wire (or
central spot within the pinhole, preferably) should have a diameter of 600 μm.
A second PMT has been added to the current system for reference purposes; it
measures light that is initially transmitted by the pellicle in the forward path. In the case
of this miniaturized system, the secondary detector could be placed in the X-Z plane after
the pellicle, with the initially light reflected up (out of the page) immediately after the
274
pellicle.
Consider the initial angle of the beam entering the system in Figure 9.7.
Another option would be to place ~2% reflectance Spectralon after the pellicle such that a
detector placed below the FSM base (or the FSWF) could monitor the raw laser light.
Figure 9.11: The schematic of the return path, with blue rays representing an illuminated
point on the retina, and the green rays representing scattered light 250 μm up and over
from the illuminated point that returns through the system. Notice that the green rays
pass through the FSWF, while the blue rays are blocked. Again, the OAPs, FSM, and
PMT are drawn to scale to indicate the space they will take up in the system.
In order to focus the collimated beam coming from the FSM in a spatially
efficient manner, another OAP is added after the pellicle to bend the light 90°; it is placed
25 mm from the pellicle to keep the system compact and allow room for the FSM base.
Again, for minimization of space and cost, OAP 5 has a BFD of 25 mm (f = 50 mm for
the paraboloid). However the best focus occurs at 20 mm from OAP 5.
275
Figure 9.11 shows the return path of both reflected and scattered light (blue and
green rays, respectively). The optical components are again drawn to scale to illustrate
the space they require in the system. The spot at the FSWF is comparable to that of the
current system, as indicated by the spot diagram in Figure 9.12. The pair of spots on the
left (from the current design) are larger, but display more chromatic variation than those
in the miniaturized system. In both cases, the top spots represent a bundle of rays coming
from light scattered and re-entering the return path from 250 μm up and 250 μm over.
The bottom spots represent the back-reflected ray bundles.
Figure 9.12: Spot diagrams of the beam returning from the eye in the current system
(left) and the miniaturized system (right). Notice that the spots in the miniaturized
system maintain more chromatic uniformity than their current counterparts due to the use
of achromatic focusing lenses. The spot sizes are listed quantitatively to the left of each
spot.
276
d. Design Process: Mounting Considerations
Because the redesigned optical system is significantly smaller and lighter than the
current system, it could be mounted to the pre-fabricated eye-imaging mount with 4
degrees of freedom that had been attached beneath the 2 x 1.5 foot optical breadboard
suspended by springs. This mount would allow for translation in the x-, y-, and zdirections, as well as rotating about the eye vertically. If this mount were then attached to
a cart on wheels, it could be rotated horizontally about the eye. However, this rotation of
the cart, or even a horizontal twisting of the mount, is not ultimately acceptable. In order
to consistently image the retina, all rotation must occur about the pupil of the patient’s
eye. All moving parts must also be able to either be locked into place or maintain their
position on their own. Ideally, the optical housing will be on a lever arm that allows for
twisting about the optical axis (of the eye) and rotation about the pupil horizontally in
addition to the motion allowed by the current mount option.
D. Further System Miniaturization
Miniaturization of the optical data acquisition system would also improve system
mobility and enable better data acquisition by making targeting more manageable, stable,
and consistent. The remainder of the system should be made small and mobile. This
should further improve targeting and data acquisition, as well as make the instrument
more versatile because of improved mobility.
One of the most important aspects of miniaturizing the system is removing the
illumination sources from the plate/board to which the rest of the optics are mounted. In
the current system, the Ar++ laser is bulky and very heavy, and the AOTF, while
relatively small, still requires several square inches. The collimation, compression, and
277
alignment of the IR diode laser with the Ar++ are also costly in terms of space. In order to
isolate these components from the rest of the optical system, it seems that fiber coupling
might be the best option. The two lasers are aligned in a similar fashion to their current
configuration: the AOTF is positioned after the Ar++ laser, while the IR laser is brought
into alignment with the AOTF-passed beam by a cold mirror.
The aligned beams will then be coupled into a fiber optic cable, which will
then deliver the light into the system as indicated in Figure 9.7. However only 75-95% of
the power incident at the fiber will actually exit the fiber, and better coupling (i.e. better
power throughput) comes at greater expense. Using the 90:10 pellicle beam splitter with
spectral dependence, the current system is such that only 25-30 μW can be delivered to
the eye in all five wavelengths, at least while relatively maintaining laser stability and a
reasonable lifetime. In order to use fiber coupling, more powerful lasers may be
required. The current laser was made by JDSU and has a total power output of ~40 mW.
While a more powerful laser will be significantly larger and heavier, it can be wheeled
around with all of the other secondary equipment that needs to be attached to but not
contained with the optical system. Similarly, a more powerful IR laser would also be
required.
If the more powerful Ar++ lasers are too bulky, multiple small but
powerful 1-wavelength lasers in the blue-green wavelengths could be co-aligned and
used instead. Regardless, the laser sources cannot be contained in the head of the ROx if
it is to be aimed with the maneuverability required for medical use.
The optical head (sans laser sources) should also be mounted on a system similar
to that of the Eye Oximeter 260, 70. The optical head was essentially mounted to a dental
278
light arm, which provides 6 degrees of freedom of motion, is easy to move, and remains
in place upon positioning. The ease of aiming is critical for the practical use of the ROx
in a clinical setting, as well as for adequate data acquisition on live subjects such as
humans and anesthetized swine.
In the current system, a gutted CPU tower is used to house the customized
electronic hardware for the ROx. While it is sufficient in production, most if not all of
the hardware could be condensed onto one or two printed circuit boards. Similarly, the
desktop computer could be replaced with a laptop. However, for data analysis,
processing speed and RAM must be considered, and a desktop may still be the better
option.
E. Conclusions
Miniaturizing the ROx is one key step to making an instrument for use in a
clinical setting. This new miniaturized design reduces the optical path by 75% and
significantly reduces the weight of the ROx head. The illumination spot size is
comparable to that of the current design, but the optics in this system are cheaper and
easier to assemble. The consolidation of the overall system is required in order to make a
clinical device.
In summary, this chapter has explored the miniaturization of the ROx. The y-y
bar diagram and Code V were used to analyze the current system as well as design a
smaller version of the optical layout. Toroidal mirrors were strategically implemented in
the new design, and further suggestions for other optical components such as lasers and
fiber coupling are also suggested. The design can probably be further improved upon,
but this optical design is an affordable starting point for the next generation ROx.
279
Chapter 10:
DISCUSSION AND CONCLUSIONS
A. Introduction
This chapter is a brief summary of what we have accomplished with the ROx
retinal oximeter, as well as challenges that still remain and suggestions as to how to
approach them. The improvements to optical layout and image quality of the ROx are
discussed in Section B, and the status and future direction of the system automation and
the automated image analysis is brought up in Section C. Recommendations for future in
vivo experiments are made in Section D, as well as a general assessment of the in vivo
experiments performed thus far. Section E includes other considerations and conclusions
that might be tied into future work.
B. Image Quality
When the ROx was first used in a live pig experiment, the noise within the image
was unacceptable, and aiming was so difficult that it was nontrivial to acquire a wellfocused image of a retinal vessel, much less identify it. Without a well-focused image, it
was also complicated to assess vessel glint and the effectiveness of the FSWF.
With implementation of the PicoScope, sampling in the fast scan direction is
increased a hundred fold, allowing for pixel binning and effectively integrating the signal.
This reduces signal spikes due to shot noise from the PMT and increases pixel bit depth,
as well. The targeting capabilities of the system were drastically improved by mounting
the optical breadboard on the jack. While the aiming is still not ideal, improvement will
280
come as the optical layout is miniaturized. There is no reason the ROx cannot be as
easily maneuverable as the EOX-238, 58.
However, with improved image quality, the presence of the glint became more
evident, as did the incorrect size and possibly placement of the FSWF.
The back
reflection from the focusing lenses also became more obvious, especially as the IR
targeting was implemented; this matter also requires correction in order to fully utilize
the entire scan region.
a. Further Glint Analysis and Prevention
The improvements to the automated analysis are evident, but it is not efficient to
further analyze the process and algorithms without better data. The resolution has been
improved, but the central glint off of vessels in a live eye is too large for the current
algorithms to handle. In addition to further analysis of the spot size and its magnification
at the FSWF, a set of experiments was performed after the analysis of data from Pig 7
(the second calibration attempt).
By a first order calculation, the magnification of a spot on the retina at the filter
plane is 6. With the current OAP mirror (OAP 5) relaying light to the filter, the
magnification is currently dependent on the focal length of the eye and the focal length of
OAP 4 (the last mirror before the focusing lenses).
Further investigation in Code V produced magnifications of 5.64 with the Arizona
eye model, and 6.07 with the eye modeled as a perfect thin lens. Examining the system
with correction for +/-5 diopters of defocus resulted in magnifications of 6.08 (myopic
eye) and 6.11 (hyperopic eye). To accomplish this, the physical length of the eye was set
to 16.78 mm, and the focal length of the perfect thin lens eye model was varied,
281
corresponding to +/- 5D. The optimization function of Code V was utilized to move the
1st focusing lens position, simulating motion of the lens via the actuator as the ROx user
finds best focus. Those distances were then used in the return path to look from a point
on the retina to the intermediate image plane between OAP4 and OAP3. The size of the
spot in this plane was compared for all three eye sizes.
In addition the magnification issue, a prevalent flaw throughout all of the later
experiments was the method of placing the FSWF. The filter was placed and adjusted by
eye: with the IR laser and AOTF off, an operator would position the filter such that the
bisecting wire was in best focus when viewed with a relaxed eye looking into the system
along the return path. The residual light (tens of nW) from the AOTF provided a bluegreen spot of light that was used to align the filter. Looking into the system, the operator
could adjust the FSWF such that the spot was centered on the wire and in the center of
the pinhole.
Instead, a more correct placement process uses a plane mirror placed in the object
plane. With the scan mirrors on and the focusing lenses collimating the beam, the mirror
is positioned such that the reflected beam exactly matches the forward path. The FSWF
is then placed after OAP 5, such that the reflected beam is precisely focused at the center
of the pinhole on the bisecting wire. The mirror is then removed and replaced with a
model eye lens (~17 mm focal length) that focuses the beam on a wire-on-paper target
(Figure 10.1). Three segments of 28 gauge wire (321 μm in diameter, similar to the
diameter of a large retinal blood vessel) were taped at different orientations to a
resolution chart showing 1 line pair per mm (Edmund Optics). The lens and target are
282
positioned as if they were the iris/lens of the eye and the vascular retina. This is an
exaggerated simulation of the reflection off of retinal vessels in vivo.
Figure 10.1: Wire-on-paper target. Segments of 28 gauge wire are taped in three
orientations to a resolution chart printed with 1 line pair per mm.
Once the test target is in place, the PMT is uncovered and the target is imaged via
live streaming with the frame grabber. The position of the FSWF is then more finely
adjusted so that the reflection from the target wires is minimized and fastened tightly in
place. This final step is necessary to compensate for any imperfection in the angle of the
plane mirror.
With this filter placement optimization better understood, the FSWF itself can be
better tested. Using the wire-on-paper target, six different filters were tested. The tested
pinhole sizes were 2.38 and 3.16 mm in diameter, corresponding to approximately 400530 μm at the retina. The tested bisecting wires had diameters of 400, 550, and 700 μm,
corresponding roughly to 67-117 μm at the retina. The plane mirror was placed in front
283
of the target and lens in the object plane in such a way that it could be pivoted in and out
of the system without moving the target and lens as each test filter is properly placed.
A preliminary set of experiments was performed, imaging the target with each
filter. A comparison is made for each filter, imaging when the reflection from the target
wires seems best blocked (reflection is on the wire of the FSWF) and when the reflection
seems to be passing through the filter uninhibited (reflection is “off” the wire of the
FSWF). It can be seen from these image that the orientation of the vessel affects the
amount of glint blocked by the filter. For instance, the horizontal and 45 o wires have a
more noticeable improvement with the filter in place than the vertical wire. This makes
sense because the ROx images are rotated 90o from the physical orientation of the target;
the horizontal line in the image is oriented the same way as the bisecting wire in the
FSWF.
This is demonstrated by comparing the photograph in Figure 10.1 to the inset
photograph in Figure 10.2. Note that both of these photographs show the approximate
region imaged by the ROx. The distortion caused by scanning is also evident.
By comparing images and cross-sections of each wire for all six test filters, the
contrast and glint reduction for each filter can be examined. Figure 10.3 shows the
difference between glint block and not glint block for each test filter. In all cases the
contrast decreases when the glint is blocked, but the wires are generally darker than the
brightest parts of the background (simulated fundus) when the filters are in place. Note
that there is no change in gain throughout the experiment, and there is no change in
object position between the adjacent images (except for the second-from-bottom right
image, where the object was slightly displaced).
284
Figure 10.2: Two sets of images acquired by the ROx of the wire-on-paper target. The
only difference between the right set of images and the left set is the lateral position of
the FSWF, demonstrating the effectiveness of the filter in blocking the glint from the
wires. The FSWF used here has a pinhole diameter of 2.38 mm bisected by a wire 550
μm in diameter. The image outlined in red (bottom center) is a photograph of the same
region of the target imaged by the ROx. The distortion introduced by scanning is clearly
evident in this comparison.
Judging from the images in Figure 10.3, the 2.38 mm pinhole filters consistently
blocked the glint best. As for bisecting wire diameter, 550 μm seems to maintain best
contrast while blocking the most glint overall (considering all orientations).
Upon looking at cross-sections of the different wires, it is evident that the 700 μm
bisecting wire in the 2.38 mm pinhole actually produces a greater glint-to-background
ratio than the 400 or 550 μm wires, probably because of the reduced contrast; the wire is
the largest fraction of the pinhole, letting less light pass overall. However, the glint could
also be affected by the different position of the wires in that particular image.
285
Figure 10.3: Images comparing all six test filters, with the reflection passing through the
filter (left) and best blocked by the filter (right). Diameters are listed to the left in units
of mm.
286
The main difference between the 400 and 550 μm wires in this data set is the
blockage of reflection from the diagonal wire. The other two wires have similar profiles,
though the glint-to-background ratio is lower on the diagonal wire and the right side of
the horizontal wire with the 550 μm filter wire. Figure 10.4 shows two filtered images
taken with the 2.38 mm pinholes: one with the 400 μm bisecting wire (A and C) and one
with the 550 μm bisecting wire (B and D). The selected regions of wire in each image
were treated as vessels, straightened and compressed into a single profile for each of the
five BGO wavelengths. The profiles of all wavelength were normalized and plotted
together for each orientation in each image.
Figure 10.4: Comparison between 400 and 550 μm bisecting wires in test FSWFs.
Pinhole diameters are 2.38 mm in both cases. The left image pair and corresponding
profiles compare the glint removal on the diagonal vessel. The right image pair and
profiles compare glint removal from similar sections of the horizontal wire in the image.
Judging from the profiles, the 550 μm wire evidently blocks more of the glint.
Notice the lower central peaks in those profiles as compared to those of the 440 μm wire.
Any variation in the profiles between wavelengths is probably due to the varying spot
sizes. Additionally, an annular filter would be expected to block only the center of the
pinhole, making the filter less sensitive to vessel orientation. The matter of optimum
filter size and shape should be studied further.
287
It has become apparent that vessels of enucleated eyes do not produce a glint like
retinal vessels in vivo (also evidenced in previous work101). This is possibly due to a
break-down and/or degradation of the inner limiting membrane (separating the vitreous
from the retina) post-mortem102, 103. Figure 10.5 shows two retinal images of enucleated
pig eyes on separate occasions. The first (top left) is the sum of all wavelengths aligned,
using best resolution with the reference signal divided out. The analysis of this wellfocused image is shown to the right. Note that this is evidence that focused images
without vessel glint can be successfully analyzed by the methods described in Chapter 8.
Figure 10.5: Two examples of retinal vessels in enucleated pig eyes. There is no vessel
glint present in these images, further exemplified by the vessel profiles (in blue) to the
right produced from the selected region in the top left image. The red line fitting the
vessel profiles is used to produce a high-quality parabolic OD spectrum.
The bottom left image was acquired after Pig 7 in attempt to align and test the
FSWF set. The eye was not of good quality, so the focus is not pristine. However, in
288
either case (or in any other enucleated eye I observed), there is no glint visible.
Enucleated eyes are useful for testing resolution and contrast, but they cannot be used to
test how well a FSWF blocks the central glint from a vessel.
b. Ghost Reflections from Lenses
The back reflections (“ghost reflections”) from the two focusing lenses can reach
the PMT, which degrades image quality, sometimes making parts of images unusable for
analysis for the reasons similar to the central glint issue. If the light returning from a
vessel or surrounding fundus is overridden by reflection, the OD cannot be determined.
Often when using the IR for targeting, the difference in reflection between IR and bluegreen (due to the limited spectrum of the AR coatings or slight alignment differences)
creates difficulties in targeting and/or image analysis; the reflection can be minimized in
one spectral range and wash out an image entirely in the other (as described in Chapter
7). In order to remove the reflections from the field of view, the lenses are often tilted
until the reflections are minimized. This is necessary for measuring vessel OD, but it
introduces aberrations into the illumination spot. This changes the spot size and the
region of the retina from which light is collected, most probably losing resolution as the
spot size and collection region on the retina are increased.
One solution to these reflections is the use of mirrors instead of lenses. A pair of
focusing mirrors could replace the focusing lenses.
In order to keep the current
orientation of the device with respect to the eye, a pair of plane folding mirrors would be
used, as well. One pair of mirrors could be mounted to an actuator, maintaining the
ability to accommodate for near- and far-sightedness. A preliminary analysis of this type
of design in Code V shows that, using two OAPs and 2 folding mirrors, aberrations
289
worsen, particularly the astigmatism as the scan angle increases. However, since the
current focusing lenses are tilted and/or decentered to avoid ghost reflections, similar or
even greater aberrations are introduced anyway.
Figure 10.6: Proposed ROx design with plane folding and OAP focusing mirrors
replacing the focusing lenses. FM1 and FOAP 1 (in the dashed box) can be translated
forward and back, indicated by the dashed arrow. Note that an afocal system with a
perfect lens is used to model the eye instead of the Arizona Eye Model.
The design illustrated in Figure 10.6 uses two 2-inch focal length OAPs (off the
shelf, Newport Optics), roughly matching the focal lengths of the achromatic lenses
currently in place. The first OAP-folding mirror pair can mounted on a single actuator,
allowing for motion required to adjust focus (indicated by the vertical dashed arrow in
Figure 10.6). The two folding mirrors could also be paired on an actuator that moves in
290
the horizontal direction (relative to the dashed arrow). Other configurations were also
briefly explored, but this is the configuration that produces the smallest spot sizes across
the illuminated region. Longer focal length OAPs are desirable to provide a longer
distance between the last pair of mirrors and the eye. This is necessary in order to
maintain maneuverability of the device during aiming.
Other design considerations
include the use of toroidal mirrors instead of OAPs for focusing optics, and if this is
useful in the current optical configuration of the ROx, it should be implemented into the
miniaturized design, as well.
C. System and Analysis Automation
In order to be a functional device, the ROx requires automation in its data
collection process. For the sake of efficiency, repeatability, and consistency, the analysis
process should be automated as much as possible.
This section summarizes the
development of these automations.
a. System Automation
The level of automation currently achieved with the ROx is functional, but there
is still much to be desired. The device triggers and collects data with a useful level of
consistency, the wavelengths and blanks switch reliably, and images are acquired, saved,
and displayed in a useful manner generally within 20-40 seconds. In fact with the
implementation of the PicoScope, the images that are displayed are more useful to the
ROx operator than before, showing the images in each wavelength individually.
The customized electronic hardware can be consolidated and stabilized (e.g. the
edge blanking could be made more consistent), but its functionality could be further
enhanced as well. For instance, the triggering signal(s) could be reworked to include
291
signal from an electrocardiogram such that every image is acquired at the same part of
the heartbeat, removing a source of variability in ROx measurements. It would be useful
to implement a laser feedback system such that the power of each wavelength is
periodically automatically optimized via controlling the AOTF.
Any effects of
microsaccades could possibly be reduced by increasing the total frame rate; this would
cut the number of data points binned per pixel in the fast-scan direction, so an analysis of
image noise at different bin values would have to be performed.
Most of the automation incorporated into the system and analysis could be
significantly streamlined and improved with dedicated support of a computer
engineer/engineering student.
The data acquisition software could be enhanced to
include more features and user-friendliness. In fact, it is conceivable to target (stream) as
well as acquire instantaneous data with the PicoScope alone, removing the need for the
frame grabber at all (and consequently the necessity of using a C++ MFC application—
the developer has many more options for programming type/language). The image
display functions created in Matlab could be replaced with functions created within the
C++ framework; passing data into a Matlab format is sometimes time-consuming. A user
interface could also be developed for interactive image analysis. This is easily developed
in Matlab, but again it would be preferable to analyze the data using the same
program/language as the data acquisition software.
Ultimately, it is desirable to have the entire image analysis and SO 2 calculation
performed and displayed immediately upon data acquisition. While some steps toward
this goal can be accomplished simply by more elegant programming (and possibly a
292
dedicated processor), the automated (or semi-automated) image analysis process is not
yet reliable enough to implement into such a system.
b. Automated Image Analysis Procedure
Chapter 8 details the current state of development of the automated image
analysis procedure.
Image quality significantly influences the effectiveness of any
analysis automation; for this discussion, it is assumed that the edge blanking, PMT gain,
and laser/AOTF power settings for each wavelength are set within normal useful ranges.
Currently, a raw dataset (from either the frame grabber or the PicoScope) can be
consistently converted into five separate images by wavelength, which can then be coaligned with each other. A vessel in the image can be regularly isolated and straightened,
though this is an area of the code that could still be improved by further specifying
conditions of a straightened vessel and its surrounding fundus. In fact, if any glint is still
present, it could be used to straighten the vessel and determine vessel edges/width91, 92.
However when it comes to determining the vessel OD, there are unresolved
complications.
A novel vessel fitting equation was implemented for use on vessel
profiles, particularly the results of averaging straightened vessels into single profiles.
The automated procedure was originally designed to incorporate this vessel fit into
analysis of the Pig 6 calibration data. At that time it was understood that though the
image quality was improved over previous experiments, it was still too noisy. It was also
clear that the calibration experiment was only roughly controlled. The animal was not
allowed to stabilize at a given SaO2 value, calling to question the true relationship
between the CO-Ox data from the femoral artery and the retinal data.
293
With these
limitations in mind, the OD analysis procedure seemed satisfactory until better images
and calibration data could be acquired.
When the PicoScope was implemented, images from enucleated pig eyes and a
healthy human eye were used to further test and modify the algorithm and fit equation.
For a well-focused image of the enucleated pig retina (which has essentially no glint), the
vessel profile produced a nice fit and a cleanly parabolic OD spectrum using the R2 value
of the parabolic fit as the metric in the iterative automated analysis process (see Figure
10.5). The relative vessel diameter and backscatter values were also determined in this
process. The repeatability for the enucleated eye image yielded OD minima differences
around 2 nm, corresponding to a difference of ~6% SO2; this is not clinically acceptable,
but it is difficult to further improve the analysis method without calibration data, which
was not available for enucleated eyes.
There were two other possible optimization metrics considered in addition to the
parabolic fit of the OD spectrum: the quality of the vessel fit itself, and the average
wavelength value over all iterations. With the evident improvement of the vessel fitting
algorithm, the R2 value of the vessel fit was deemed a reasonable indicator of analysis
fidelity. This led to the inclusion of the R2 value of the vessel fit as an optimization
metric. The upgraded analysis seemed to improve the quality of the calibration line.
However, its repeatability is simply not precise enough, and the questionable accuracy of
the calibration experiment makes it difficult to judge the accuracy of the ROx data,
though there are apparent issues with the offset of the ROx data, perhaps due to poor
alignment of the FSWF.
294
The data from the Pig 7 experiment proved to have too much glint for the
automation to handle. The algorithm could not remove the glint enough to make reliable
fits to the vessel profile, and consequently the OD spectra are not dependable. The glint
is so significant in most images that even vessel profiles fit by hand would be
questionable.
In conclusion, the automation of the data analysis is certainly progress. However,
the analysis procedure has much left to be desired. The repeatability of analysis on a
single image is not consistent and worsens when different people analyze an image. The
uncertainty and error, while highly dependent on image quality, are substandard. The
analysis is good enough for an experienced user to find trends, but it is currently
impossible to use it for a dependable quantitative analysis of retinal SO2.
There are many areas of the algorithm that should be revisited and further
developed. Many of the boundaries imposed throughout the algorithm are empirically
based and should be better defined and/or modified. Among the factors needing better
definitions are fundus values. The spectral neutrality of the fundus has been discussed,
but some regions of the fundus were found to be more spectrally neutral than others when
comparing the 5 images from any given data set collected by the ROx. How spectrally
neutral is “neutral enough”? Several studies have shown the fundus to be spectrally
neutral enough for use as the reference in retinal OD calculations 20, 36, 50, 104, but other
studies have documented a nontrivial effect of fundus pigmentation in retinal SO2
calculations51. When determining how to define the fundus in the automated algorithm, a
potentially significant difference was observed between the OD spectrum calculated
using the “spectrally neutral” fundus versus the entire normalized fundus in the avascular
295
region of a vessel profile. Other aspects of the analysis (particularly omitting the glint
and fitting the vessel profile) took priority, but it is very important to study these
differences in order to best define the fundus used to calculated vessel OD in future work.
Fundus analysis considerations should also be made when the ROx is used for human
testing and diagnostics.
Additional work is also required in better defining the glint in a vessel in order to
best omit it. While it is critical to physically block more of the specular reflection in the
images, it is important for any residual glint to be omitted in post-processing. This
should improve vessel fitting, and in turn, improve the quality of the OD spectrum and
SO2 calculations.
The optimization metric should also be further explored. If the glint is removed,
the dual-metric process may prove unnecessary (as with the enucleated eye), and a single
metric could be used to determine the SO2 calculation. In addition to the R2 values the
vessel and OD spectrum fits, it is conceivable to use the average value of the wavelength
of the OD minimum over all iterations (perhaps omitting outliers). This method was not
truly tested, but during the development of the dual-metric system, it was observed as a
plausible alternative.
c. Other Considerations
There are potentially more aspects of the ROx system and retinal physiology that
could be considered and incorporated into the automation algorithm, not the least of
which are the relationships between retinal SO2, the vessel diameter, and scattering in the
blood, particularly in the blue-green wavelengths.
296
There have recently been studies linking SO2 to vessel diameter51, 63. This analysis
is already measuring the relative vessel diameter; if the measurement calculations were to
be improved and the scan angle or image size were to be more precisely measured, the
vessel diameter could be incorporated as a factor in the final SO2 calculation. Accurate
vessel diameter measurements are also useful in determining other physiological values,
such as retinal vasoconstriction/dilation and blood flow91.
The scattering within the retina, particularly backscatter from blood, has a
nontrivial spectral signature, affecting the measured OD spectrum of retinal vessels,
subsequently affecting SO2 calculations. The current analysis assumes that the scattering
term can be considered a linear contribution to the parabolic form of the OD spectrum.
These assumptions are based on prior work observing the OD spectrum of hemoglobin
and whole blood in both transmission and reflection, where the scattering component
shifted the OD spectrum of whole blood towards lower wavelengths in reflection (using
off-axis intravitreal illumination in vivo)19, 20.
It should be noted that those reflection measurements were made using off-axis
intravitreal illumination, effectively allowing isolation of the spectrum in reflection. The
ROx is an on-axis illumination system, so the light paths detected by the PMT are more
complicated. In addition to the specular reflection, some of the light is backscattered,
resulting in the reflection spectrum.
However, if light is scattered and transmitted
through the vessel and reflected by the fundus into the collection path, that light would
contribute to the spectrum in transmission. This adds variables to the BGO technique
that have not yet been explored.
297
However, some of the on-axis in vivo data analysis indicates a shift toward higher
wavelengths by comparison to the OD spectrum of Hb. One possible explanation is that,
due to the FSWF, the reflection/backscattered path is blocked and light that has been
scattered and/or transmitted by the vessel (single-pass light) has reflected off of the
surrounding fundus and returned to the detector; the FSWF could be effectively blocking
the spectral data in reflection but collecting the spectral data in transmission that has been
reflected/scattered by the fundus. In that case, the edge of the vessel profile would
possibly consist of a combination of backscattered light and single-pass light that has
reflected/scattered off of the fundus and back up into the return path through the vessel.
This idea is actually supported in a recent study by Rodmell et al.
The
contribution of double-pass light has been shown to be negligible, and the path length of
light interacting with a retinal vessel is shown to be heavily influenced by the vessel
diameter. In fact, it is shown that with offset illumination and detection paths (e.g.
intentionally placing the FSWF off-axis, or even just blocking enough of the central
illumination as described in Section C.a), the path length is very nearly the vessel
diameter, enough so that it could be considered constant for all collected light from that
vessel. The offset between illumination and detection paths effectively ensures that
single-pass light is isolated from double-pass and backscattered photons.
If the FSWF is not properly aligned and/or collects light from a very small spot on
the retina, the reverse could conceivably be observed: the backscattered light (as well as
the glint) would be detected while more of the single-pass light is rejected. In the cases
in which the alignment of the FSWF is in question (including the one mentioned above),
the contribution of backscattered and single-pass light could both be present. The Pig 6
298
calibration data seems to have a particularly large shift, larger than that of the isolated
whole blood in transmission19. Perhaps the globe of the eye acting as an integrating
sphere increased the effect of scattering in transmission such that the spectral shift was
larger than before, where no integrating sphere appears to have been used.
With improved image quality and a better grasp of the requirements for the
FSWF, another calibration experiment should be performed with the expectation of
determining a believable calibration line. However, the effect of the position of the
FSWF on the offset of the calibration line should be studied further.
A model or
enucleated eye could be used at first, but it would be ideal to be able to take multiple sets
of calibration data on a live pig eye, each with the FSWF in a different
position/configuration. Further incorporation of the vessel diameter into the vessel fit
and OD calculations should also be considered.
Additionally, it might be interesting to look at improving the confocality of the
ROx by using a smaller pinhole (~2 mm) with no bisecting wire, but set the pinhole offaxis to collect scattered light and reject the backscattered light. This would probably
cause the images to be darker, but this might be a solution if consistently tighter focus is
desired.
D. In Vivo Experiments
The progress of the ROx development is seen in its effectiveness during in vivo
experiments: it was used for calibration, sepsis, and human eye experiments.
The calibration experiments produced rough calibration lines, but more
experiments are required with the ROx updated, specifically with the appropriate FSWF
properly aligned. At least 4 good calibration experiments should be performed in order
299
to match the present standard20, and neither of the previous experiments should be
contributed to that count.
The calibration should be performed before any other
physiological change is introduced (e.g. septic insult), inspired O2 should be controllable
such that the animal is allowed to stabilize at a given SaO2, and ROx measurements of a
retinal artery should be made simultaneously with blood draw from the femoral artery to
be measured via the CO-Ox.
The sepsis experiments did not produce any quantitative SO2 data, but general
trends were observed between SrvO2 and SvO2 (and to a slightly lesser degree, ScvO2).
They also verified a novel model for quickly and consistently inducing controllable but
severe sepsis and septic shock in pigs. As future experiments continue, ROx data should
be taken as often as vital signs are recorded.
This will help relate SrvO2 to other
physiologic responses during the progression of sepsis. If possible, an artery-vein pair
should be imaged in order to measure the oxygen extraction ratio of the eye, as well. The
eye should also be paralyzed in future pig experiments until a smaller, more
maneuverable optical head is built. As the automated analysis is improved and the
behavior of SrvO2 during the onset of sepsis is better understood, the ROx should be used
to guide the resuscitation of the animal and potentially test new treatments. Other nearfuture swine experiments could include tracking SrvO2 over the onset of other shock
states, particularly neurogenic shock, occurring when the spinal cord is severed.
The ROx acquired multiple images on multiple occasions from the eye of a
healthy human volunteer, but a satisfactory calibrated measurement was never achieved,
and the image quality remained inconsistent. Many changes were made to the ROx to
improve patient comfort, image quality, and operator convenience.
300
However, it is
particularly necessary for these experiments that the AOTF/IR alignment be improved
and stabilized, and that the focusing lenses be replaced with mirrors, removing the ghost
reflections from both targeting and image acquisition modes. The blue-green and IR
images need to be as similar as possible in terms of quality in order to consistently collect
useful images from a human patient.
Many trends are being generated, relating human retinal SO2 to age, gender,
left/right eye, chronic obstructive pulmonary disease (COPD), diabetic retinopathy, and
retinal vein occlusion to name a few90, 95, 105-108. If the ROx is developed into a clinically
accurate device, it might be interesting to verify some of these findings with more
quantitative accuracy.
E. Summary
The ROx retinal oximeter has been developed to a point of usefulness in acquiring
data from humans and anesthetized pigs in vivo.
It has not yet been successfully
calibrated, though improvements have been made to that end. It has been used to roughly
assess the correlation between SrvO2, SvO2, and ScvO2 during the onset and development
of severe sepsis and septic shock. The ROx has been modified to accommodate imaging
a human eye, though there is still much room for improvement. An automated analysis
algorithm is under development, but the quality of analysis varies considerably with
image quality. The analysis procedure warrants further scrutiny and consideration as the
vessel fit equation and optimization metrics are improved upon. A miniaturized, more
maneuverable optical layout has been designed and presented. The ROx is a prototype
with great potential to be used as a ground-breaking diagnostic in future medical
research.
301
REFERENCES
1. Medical Physiology: A Cellular and Molectular Approach, W. F. Boron and E. L.
Boulpaep, Eds., 2nd Saunders, Philadelphia, (2009).
2. J. SchwiegerlingField Guide to Visual and Ophthalmic Optics, Anonymous , Eds., pp.
109, SPIE Publications, (2004).
3. O. W. van AssendelftSpectrophotometry of Haemoglobin Derivatives, C. C. Smith,
Eds., Springfield, IL, (1970).
4. M. Meinke, G. Mueller, J. Helfmann, M. Friebel, "Optical properties of platelets and
blood plasma and their influence on the optical behavior of whole blood in the visible to
near infrared wavelength range," J Biomed Opt 12(1), 014024 (2007).
5. Salyer DA. Fundus Spectroscopy and Studies in Retinal Oximetry using Intravitreal
Illumination. [PhD]. University of Arizona; 2006.
6. R. A. Abrams, D. E. Meyer, and S. Kornblum, "Speed and accuracy of saccadic eye
movements: Characteristics of impulse variability and the oculomotor system," Journal
of Experimental Psychology: Human Perception and Performance 15(3), 529-543
(1989).
7. I-STAT |Blood Analysis | Handheld Meter | Abbott Point of Care.com Available from
URL: http://www.abbottpointofcare.com/Patient-Care-Settings/Hospital/CriticalCare.aspx [accessed 10/29/2014, 2014].
8. J. Severinghaus, P. Astrup, and J. Murray, "Blood gas analysis and critical care
medicine," American Journal of Respiratory and Critical Care Medicine 157(4), S114S122 (1998).
9. K. Kramer, J. Elam, G. Saxton, and J. W. Elam, "Influence of oxygen saturation,
erythrocyte concentration and optical depth upon the red and near-infrared light
transmittance of whole blood," American Journal of Physiology 165229-246 (1951).
10. G. Millikan, "The oximeter, an instrument for measuring continuously the oxygen
saturation of arterial blood in man," Rev Sci Instrum 13(10), 434-444 (1942).
11. J. Severinghaus, "Theory and history of oximetry," Acta Anaesthesiol Scand 3181-81
(1987).
12. T. Aoyagi, M. Kishi, K. Yamaguchi, and S. Watanabe, "Improvement of the earpiece
oximeter," Abstracts of the 13th annual meeting of the Japanese Society of Medical
Electronics and Biological Engineering 90-91 (1974).
302
13. A. Zwart, A. Buursma, B. Oeseburg, and W. Zijlstra, "Determination of hemoglobin
derivatives with the il-282 co-oximeter as compared with a manual spectrophotometric 5wavelength method," Clin Chem 27(11), 1903-1907 (1981).
14. V. Twersky, "Multiple scattering of waves and optical phenomena " J Opt Soc Am
52(2), 145 (1962).
15. V. Twersky, "Absorption and multiple scattering by biological suspensions," J Opt
Soc Am 60(8), 1084-& (1970).
16. A. J. Cohen and R. A. Laing, "Multiple-scattering analysis of retinal blood oximetry,"
IEEE Transactions on Biomedical Engineering 23(5), 391-400 (1976).
17. J. M. Steinke and A. P. Shepherd, "Diffusion model of the optical absorbance of
whole blood," Journal of the Optical Society of America A 5(6), 813-822 (1988).
18. J. M. Steinke and A. P. Shepherd, "Comparison of mie theory and the light-scattering
of red blood-cells," Appl Opt 27(19), 4027-4033 (1988).
19. K. R. Denninghoff, R. A. Chipman, L. W. Hillman, "Blood oxyhemoglobin saturation
measurements by blue-green spectral shift," Journal of Biomedical Optics 12(3), 034020
(2007).
20. K. R. Denninghoff, D. A. Salyer, S. Basavanthappa, R. I. Park, R. A. Chipman,
"Blue-green spectral minimum correlates with oxyhemoglobin saturation in vivo," J
Biomed Opt 13(5), 054059 (2008).
21. D. J. Faber, E. G. Mik, M. C. G. Aalders, T. G. van Leeuwen, "Light absorption of
(oxy-)hemoglobin assessed by spectroscopic optical coherence tomography," Optics
Letters 28(16), 1436-1438 (2003).
22. L. Kagemann, G. Wollstein, M. Wojtkowski et al., "Spectral oximetry assessed with
high-speed ultra-high-resolution optical coherence tomography," J Biomed Opt 12(4),
041212 (2007).
23. G. Lu and B. Fei, "Medical hyperspectral imaging: A review," J Biomed Opt 19(1),
010901 (2014).
24. B. Khoobehi, J. Beach, and H. Kawano, "Hyperspectral imaging for measurement of
oxygen saturation in the optic nerve head," Invest Ophthalmol Vis Sci 45(5), 1464-1472
(2004).
25. W. R. Johnson, D. W. Wilson, W. Fink, M. Humayun, and G. Bearman, "Snapshot
hyperspectral imaging in ophthalmology," J Biomed Opt 12(1), 014036 (2007).
303
26. S. R. Patel, J. G. Flanagan, A. M. Shahidi, J. Sylvestre, and C. Hudson, "A prototype
hyperspectral system with a tunable laser source for retinal vessel imaging," Investigative
Ophthalmology & Visual Science} 54(8), 5163-5168 (2013).
27. J. B. Hickam, R. Frayser, J. C. Ross, "A study of retinal venous blood oxygen
saturation in human subjects by photographic means," Circulation 27(3), 375-385 (1963).
28. S. H. Hardarson, A. Harris, R. A. Karlsson, G. H. Halldorsson, L. Kagemann, E.
Rechtman, G. M. Zoega, T. Eysteinsson, J. A. Benediktsson, A. Thorsteinsson, P. K.
Jensen, J. Beach, E. Stefansson, "Automatic retinal oximetry," Investigative
ophthalmology & visual science 47(11), 5011-5016 (2006).
29. D. Schweitzer, L. Leistritz, M. Hammer, M. Scibor, U. Bartsch, J. Strobel,
"Calibration-free measurement of the oxygen saturation in human retinal vessels," Proc
SPIE 2393(1), 210-218 (1995).
30. D. J. Mordant, I. Al-Abboud, G. Muyo et al., "Validation of human whole blood
oximetry, using a hyperspectral fundus camera with a model eye," Invest Ophthalmol Vis
Sci 52(5), 2851-2859 (2011).
31. D. J. Mordant, I. Al-Abboud, G. Muyo et al., "Spectral imaging of the retina," Eye
25(3), 309-320 (2011).
32. F. Reinholz, R. A. Ashman, and R. H. Eikelboom, "Simultaneous three wavelength
imaging with a scanning laser ophthalmoscope," Cytometry 37(3), 165-170 (1999).
33. R. A. Ashman, F. Reinholz, and R. H. Eikelboom, "Oximetry with a multiple
wavelength SLO," Int Ophthalmol 23(4-6), 343-346 (2001).
34. M. Hammer, W. Vilser, T. Riemer, D. Schweitzer, "Retinal vessel oximetrycalibration, compensation for vessel diameter and fundus pigmentation, and
reproducibility," J Biomed Opt 13(5), 054015 (2008).
35. R. E. K. Man, R. Kawasaki, Z. Wu et al., "Reliability and reproducibility of retinal
oxygen saturation measurements using a predefined peri-papillary annulus," Acta
Ophthalmol 91(8), E590-E594 (2013).
36. F. C. Delori, "Noninvasive technique for oximetry of blood in retinal vessels," Appl
Opt 27(6), 1113-1125 (1988).
37. L. C. Heaton, M. H. Smith, K. R. Denninghoff, L. W. Hillman, "Handheld fourwavelength retinal vessel oximeter," Ophthalmic Technologies X 3908(1), 227-233
(2000).
38. Smith MH. Oximetry of Blood in Retinal Vessels. [Ph.D]. Ann Arbor, MI: The
University of Alabama in Huntsville; 1996.
304
39. A. Lompado, M. H. Smith, L. W. Hillman, K. R. Denninghoff, "Multispectral
confocal scanning laser ophthalmoscope for retinal vessel oximetry," Proc SPIE 3920(1),
67-73 (2000).
40. A. E. Elsner, S. A. Burns, G. W. Hughes, and R. H. Webb, "Reflectometry with a
scanning laser ophthalmoscope," Appl Opt 31(19), 3697-3710 (1992).
41. H. Li, J. Lu, G. Shi, and Y. Zhang, "Measurement of oxygen saturation in small
retinal vessels with adaptive optics confocal scanning laser ophthalmoscope," J Biomed
Opt 16(11), 110504 (2011).
42. S. H. Rasta, A. Manivannan, and P. F. Sharp, "Spectral imaging technique for retinal
perfusion detection using confocal scanning laser ophthalmoscopy," J Biomed Opt
17(11), 116005 (2012).
43. R. N. Pittman and B. R. Duling, "New method for measurement of percent
oxyhemoglobin," Journal of applied physiology 38(2), 315-320 (1975).
44. M. H. Smith, "Optimum wavelength combinations for retinal vessel oximetry,"
Applied Optics 38(1), 258-267 (1999).
45. Prahl S. Tabulated Molar Extinction Coefficient for Hemoglobin in Water. Available
from URL: http://omlc.org/spectra/hemoglobin/summary.html [accessed 10/19, 2014].
46. Denninghoff KR, Hillman LW, Hillman S, inventorsMethod and Device for
Determining Oxygen Saturation of Hemoglobin, for Determining Hematocrit of Blood,
and/or for Detecting Macular Degeneration. United States 2010.
47. K. R. Denninghoff, K. B. Sieluzycka, J. K. Hendryx, T. J. Ririe, L. DeLuca, and R. A.
Chipman, "Retinal oximeter for the blue-green oximetry technique," J Biomed Opt
16(10), 107004 (2011).
48. M. H. Smith, K. R. Denninghoff, A. Lompado, L. W. Hillman, "Effect of multiple
light paths on retinal vessel oximetry," Applied Optics 39(7), 1183-1193 (2000).
49. P. I. Rodmell, J. A. Crowe, A. Gorman et al., "Light path-length distributions within
the retina," J Biomed Opt 19(3), 036008 (2014).
50. D. A. Salyer, K. R. Denninghoff, N. Beaudry, S. Basavanthappa, R. I. Park, R. A.
Chipman, "Diffuse spectral fundus reflectance measured using subretinally placed
spectralon," J Biomed Opt 13(4), 044004 (2008).
51. M. Hammer, T. Riemer, W. Vilser, S. Gehlert, D. Schweitzer, "A new imaging
technique for retinal vessel oximetry: Principles and first clinical results in patients with
retinal arterial occlusion and diabetic retinopathy," Ophthalmic Technologies XIX
7163(1), 71630P (2009).
305
52. F. C. Delori and K. P. Pflibsen, "Spectral reflectance of the human ocular fundus,"
Appl Opt 28(6), 1061-1077 (1989).
53. D. Salyer, N. Beaudry, S. Basavanthappa et al., "Retinal oximetry using intravitreal
illumination," Curr Eye Res 31(7), 617-627 (2006).
54. P. N. Youssef, N. Sheibani, and D. M. Albert, "Retinal light toxicity " Eye (Lond)
(2010).
55. F. C. Delori, R. H. Webb, and D. H. Sliney, "Maximum permissible exposures for
ocular safety (ANSI 2000), with emphasis on ophthalmic devices," J Opt Soc Am A
24(5), 1250-1265 (2007).
56. ThorLabs, "Pellicle beamsplitter," .
57. Sieluzycka KB. The Design and Testing of a New Retinal Oximeter. [MS]. University
of Arizona; 2011.
58. Lompado A. A Confocal Scanning Laser Ophthalmoscope for Retinal Vessel
Oximetry. [PhD]. University of Alabama in Huntsville; 1999.
59. L. Deschaepdrijver, P. Simoens, L. Pollet, H. Lauwers, and J. J. Delaey,
"Morphological and clinical-study of the retinal circulation in the miniature pig .B.
fluorescein angiography of the retina," Exp Eye Res 54(6), 975-985 (1992).
60. K. R. Denninghoff, M. H. Smith, R. A. Chipman, L. W. Hillman, P. M. Jester, C. E.
Hughes, F. Kuhn, L. W. Rue, "Retinal large vessel oxygen saturations correlate with
early blood loss and hypoxia in anesthetized swine," Journal of Trauma-Injury Infection
and Critical Care 43(1), 29-34 (1997).
61. S. Palkovits, M. Lasta, R. Told et al., "Retinal oxygen metabolism during normoxia
and hyperoxia in healthy subjects " Invest Ophthalmol Vis Sci 55(8), 4707-4713 (2014).
62. A. Luksch, G. Garhöfer, A. Imhof et al., "Effect of inhalation of different mixtures of
O2 and CO2 on retinal blood flow," British Journal of Ophthalmology} 86(10), 11431147 (2002).
63. J. M. Beach, A. Harris, B. A. Siesky, Y. Arieli, A. Pickrell, "Longitudinal oxygen
gradients affect corrections for vessel diameter sensitivity in retinal oximetry," ARVO
(2012).
64. K. S. Canavan, A. Dark, and M. A. Garrioch, "Sub‐Tenon’s administration of local
anaesthetic: A review of the technique," British Journal of Anaesthesia} 90(6), 787-793
(2003).
306
65. J. Ahn, M. Jeong, Y. Park et al., "Comparison of systemic atracurium, retrobulbar
lidocaine, and sub-tenon's lidocaine injections in akinesia and mydriasis in dogs," Vet
Ophthalmol 16(6), 440-445 (2013).
66. R. P. Dellinger, M. M. Levy, and J. M. Carlet, "Surviving sepsis campaign:
International guidelines for management of severe sepsis and septic shock: 2008 (vol 36,
pg 296, 2008)," Crit Care Med 36(4), 1394-1396 (2008).
67. A. E. Jones and M. A. Puskarich, "Sepsis-induced tissue hypoperfusion," Crit Care
Clin 25(4), 769-+ (2009).
68. E. Rivers, B. Nguyen, S. Havstad, J. Ressler, A. Muzzin, B. Knoblich, E. Peterson,
M. Tomlanovich, "Early goal-directed therapy in the treatment of severe sepsis and septic
shock," New England Journal of Medicine 345(19), 1368-1377 (2001).
69. L. e. a. DeLuca, "Comparison of central and peripheral venous pressures in critically
ill patients: A pilot study," University of Arizona: 2009 AHSC Frontiers in Biomedical
Research Poster Forum, Tucson AZ 1 (2009).
70. M. H. Smith, K. R. Denninghoff, L. W. Hillman, and R. A. Chipman, "Oxygen
saturation measurements of blood in retinal vessels during blood loss," J Biomed Opt
3(3), 296-303 (1998).
71. K. R. Denninghoff, M. H. Smith, A. Lompado, L. W. Hillman, "Retinal venous
oxygen saturation and cardiac output during controlled hemorrhage and resuscitation,"
Journal of Applied Physiology 94(3), 891-896 (2003).
72. D. Rittirsch, M. S. Huber-Lang, M. A. Flierl, and P. A. Ward, "Immunodesign of
experimental sepsis by cecal ligation and puncture " Nat Protoc 4(1), 31-36 (2009).
73. K. Kazarian, P. Perdue, W. Lynch et al., "Porcine peritoneal sepsis - modeling for
clinical relevance," Shock 1(3), 201-212 (1994).
74. C. C. Moore, I. H. McKillop, and Toan Huynh, "MicroRNA expression following
activated protein C treatment during septic shock," J Surg Res 182(1), 116-126 (2013).
75. K. A. Wichterman, A. E. Baue, and I. H. Chaudry, "Sepsis and septic shock—A
review of laboratory models and a proposal," J Surg Res 29(2), 189-201 (1980).
76. M. Browne and G. Leslie, "Animal-models of peritonitis," Surgery Gynecology &
Obstetrics 143(5), 738-740 (1976).
77. A. Thal, R. Robinson, T. Nagamine, R. Pruett, and A. Wegst, "Critical relationship of
intravascular blood-volume and vascular capacitance in sepsis," Surgery Gynecology &
Obstetrics 143(1), 17-22 (1976).
307
78. M. Imamura and G. Clowes, "Hepatic blood-flow and oxygen-consumption in
starvation, sepsis and septic shock," Surgery Gynecology & Obstetrics 141(1), 27-34
(1975).
79. M. Rady, "Bench-to-bedside review: Resuscitation in the emergency department,"
Critical Care 9(2), 170-176 (2005).
80. R. J. Beale, S. M. Hollenberg, J. L. Vincent, and J. E. Parrillo, "Vasopressor and
inotropic support in septic shock: An evidence-based review," Crit Care Med 32(11
Suppl), S455-65 (2004).
81. J. L. Vincent and H. Gerlach, "Fluid resuscitation in severe sepsis and septic shock:
An evidence-based review," Crit Care Med 32(11 Suppl), S451-4 (2004).
82. D. Schweitzer, A. Lasch, S. van der Vorst et al., "Change of retinal oxygen saturation
in healthy subjects and in early stages of diabetic retinopathy during breathing of 100 %
oxygen," Klin Monatsbl Augenheilkd 224(5), 402-410 (2007).
83. L. Gattinoni, L. Brazzi, P. Pelosi et al., "A trial of goal-oriented hemodynamic
therapy in critically ill patients," N Engl J Med 333(16), 1025-1032 (1995).
84. R. Dellinger, J. Carlet, H. Masur et al., "Surviving sepsis campaign guidelines for
management of severe sepsis and septic shock," Crit Care Med 32(3), 858-873 (2004).
85. K. Reinhart, T. Rudolph, D. L. Bredle, L. Hannemann, and S. M. Cain, "Comparison
of central-venous to mixed-venous oxygen-saturation during changes in oxygen-supply
demand," Chest 95(6), 1216-1221 (1989).
86. F. Bloos and K. Reinhart, "Venous oximetry," Intensive Care Med 31(7), 911-913
(2005).
87. J. D. Edwards and R. M. Mayall, "Importance of the sampling site for measurement
of mixed venous oxygen saturation in shock," Crit Care Med 26(8), 1356-1360 (1998).
88. A. MeierHellmann, M. Specht, L. Hannemann, H. Hassel, D. L. Bredle, and K.
Reinhart, "Splanchnic blood flow is greater in septic shock treated with norepinephrine
than in severe sepsis," Intensive Care Med 22(12), 1354-1359 (1996).
89. P. van Beest, G. Wietasch, T. Scheeren, P. Spronk, and M. Kuiper, "Clinical review:
Use of venous oxygen saturations as a goal - a yet unfinished puzzle," Critical Care
15(5), 232 (2011).
90. R. E. Man, M. B. Sasongko, R. Kawasaki et al., "Associations of retinal oximetry in
healthy young adults " Invest Ophthalmol Vis Sci 55(3), 1763-1769 (2014).
308
91. M. O’ Halloran, E. O’Donoghue, and C. Dainty, "Measurement of the retinal
arteriolar response to a hyperoxic provocation in nonsmokers and smokers, using a highresolution confocal scanning laser ophthalmoscope," J Biomed Opt 19(7), 076012 (2014).
92. N. Chapman, "Computer algorithms for the automated measurement of retinal
arteriolar diameters " Br J Ophthalmol 85(1), 74 <last_page> 79 (2001).
93. T. J. Ririe, "Vessel profile equation," (2013).
94. Gerstman B. StatPrimer. Available from URL:
http://www.sjsu.edu/faculty/gerstman/StatPrimer/ [accessed 10/29/2014, 2014].
95. M. Hammer, W. Vilser, T. Riemer, A. Mandecka, D. Schweitzer, U. Kuehn, J.
Dawczynski, F. Liemt, J. Strobel, "Diabetic patients with retinopathy show increased
retinal venous oxygen saturation," Graefes Archive for Clinical and Experimental
Ophthalmology 247(8), 1025-1030 (2009).
96. J. E. GreivenkampField Guide to Geometrical Optics, Anonymous , Eds., SPIE
Publications, (2004).
97. E. Delano, "First-order design and the y, y¯ diagram," Appl Opt 2(12), 1251-1256
(1963).
98. D. Malacara-Hernandez and Z. MalacaraHandbook of Optical Design, Anonymous ,
Eds., pp. 522, 2nd Marcel Dekker, Inc., New York, NY, (2004).
99. D. Malacara-Hernandez, "Some parameters and characteristics of an off-axis
paraboloid " Optical Engineering 30(9), 1277 (1991).
100. M. Born and E. WolfPrinciples of Optics, Anonymous , Eds., 6th Pergamon Press,
New York, (1989).
101. D. A. Salyer, K. Twietmeyer, N. Beaudry, S. Basavanthappa, R. I. Park, and R.
Chipman, "In vitro multispectral diffuse reflectance measurements of the porcine
fundus," Invest Ophthalmol Vis Sci 46(6), 2120-2124 (2005).
102. F. Delori, E. Gragoudas, R. Francisco, and R. Pruett, "Monochromatic
ophthalmoscopy and fundus photography - normal fundus," Arch Ophthalmol 95(5), 861868 (1977).
103. M. H. Smith, K. R. Denninghoff, A. Lompado, L. W. Hillman, "Retinal vessel
oximetry: Toward absolute calibration," Ophthalmic Technologies X 3908(1), 217-226
(2000).
309
104. J. Beach, K. Schwenzer, S. Srinivas, D. Kim, and J. Tiedeman, "Oximetry of retinal
vessels by dual-wavelength imaging: Calibration and influence of pigmentation," J Appl
Physiol 86(2), 748-758 (1999).
105. S. Palkovits, M. Lasta, A. Boltz et al., "Measurement of retinal oxygen saturation in
patients with chronic obstructive pulmonary disease " Invest Ophthalmol Vis Sci 54(2),
1008-1013 (2013).
106. A. Geirsdottir, S. H. Hardarson, O. Palsson, O. B. Olafsdottir, and E. Stefansson,
"Retinal oxygen metabolism is affected in age-related macular degeneration," Invest
Ophthalmol Vis Sci 53(6), 5178 (2012).
107. S. H. Hardarson and E. Stefansson, "Retinal oximetry in diabetes and vein
occlusions," Acta Ophthalmol 88141 (2010).
108. S. H. Hardarson and E. Stefansson, "Oxygen saturation in branch retinal vein
occlusion," Acta Ophthalmol 90(5), 466-470 (2012).
310
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising