Clinical Optical Coherence Tomography

Angiography Registration and Analysis

by

Morgan Lindsay Heisler

B.A.Sc. (Hons.), Simon Fraser University, 2015

Thesis Submitted in Partial Fulfillment of the

Requirements for the Degree of

Master of Applied Science

in the

School of Engineering Science

Faculty of Applied Sciences

© Morgan Lindsay Heisler

SIMON FRASER UNIVERSITY

Spring 2017

Copyright in this work rests with the author. Please ensure that any reproduction or re-use is done in accordance with the relevant national copyright legislation.

ii

Approval

Name: Morgan Lindsay Heisler

Degree: Master of Applied Science

Title: Clinical Optical Coherence Tomography

Angiography Registration and Analysis

Examining Committee: Chair: Dr. Ash M. Parameswaran, P. Eng. Professor

Dr. Marinko V. Sarunic, P. Eng. Senior Supervisor Professor

Dr. Mirza Faisal Beg, P. Eng. Supervisor Professor

Dr. Yifan Jian Supervisor Adjunct Professor

Dr. Paul J. Mackenzie Supervisor Clinical Assistant Professor Ophthalmology and Visual Sciences University of British Columbia

Dr. Eduardo V. Navajas External Examiner Clinical Assistant Professor Ophthalmology and Visual Sciences University of British Columbia

Date Defended/Approved: April 20th, 2017

iii

Ethics Statement

iv

Abstract

Optical Coherence Tomography Angiography (OCT-A) is an emerging imaging modality

with which the retinal circulation can be visualized by computing the decorrelation signal

on a pixel-by-pixel basis. This non-invasive, in vivo visualization of the retinal

microvasculature can be instrumental in studying the onset and development of retinal

vascular diseases. Quantitative measurements, such as capillary density, can be used to

stratify the risk of disease progression, visual loss, and also for monitoring the course of

disease. Due to projection artifact and poor contrast, it is often difficult to trace individual

vessels when only one en face image is visualized. Averaging of up to 10 serially

acquired OCT-A images with parallel strip-wise microsaccadic noise removal and

localized nonrigid registration is presented. Additionally, the use of a deep learning

method for the quantification of Foveal Avascular Zone (FAZ) parameters and perifoveal

capillary density of prototype and commercial OCT-A platforms in both healthy and

diabetic eyes is evaluated.

Keywords: optical coherence tomography; ophthalmology; image processing;

registration; retina; angiography

v

Acknowledgements

I would like to express my deepest gratitude to my senior supervisor, Dr.

Marinko V. Sarunic, whose expertise, understanding, and patience added

considerably to my graduate experience. I appreciate his confidence and trust in my

abilities while pushing me outside my academic comfort zone.

I would also like to sincerely thank Dr. Mirza Faisal Beg for all of his expertise

in medical image processing which was crucial to the completion of this research.

His passion and enthusiasm for the research made every interaction an enjoyable

experience.

I would also like to acknowledge the medical professionals who have helped

shape my graduate career. Dr. Paul J. Mackenzie and Dr. Eduardo V. Navajas both

took time out of their busy clinical practices to share invaluable medical expertise

and direction for these projects. A special thanks also goes out to Dr. Zaid Mammo

for his dedication to the research. His patience and willingness to explain concepts

that must have seemed very basic to a resident of his caliber were so appreciated.

Additionally, I would like to thank the international collaborators who have

helped with this research. Dr. Dao-Yi Yu and Dr. Chandrakumar Balaratnasingam

were extremely welcoming and supportive during my time in Perth as well as

everyone else in the Pathology and Physiology Department at Lions Eye Institute.

Also, a huge thank you to Dr. Pavle Prentašić for exposing me to interesting world of

machine learning.

I am very grateful to be part of the Biomedical Optics Research Group

(BORG) at SFU. I am especially grateful to BORG members Dr. MyeongJin Ju, Dr.

Sieun Lee and Dr. Yifan Jian for their mentorship in these past few years. They have

demonstrated some of the traits that I can only strive for in future endeavours: Dr.

Ju’s dedicated work ethic, Dr. Lee’s eye for detail, and Dr. Jian’s willingness to make

the work environment as enjoyable as possible.

Lastly, I would like to thank my family for their support and encouragement in

all my endeavours. I couldn’t have done it without you.

vi

Table of Contents

Approval ............................................................................................................................ ii

Ethics Statement ............................................................................................................... iii

Abstract ............................................................................................................................ iv

Acknowledgements ........................................................................................................... v

Table of Contents ............................................................................................................. vi

List of Tables ................................................................................................................... viii

List of Figures................................................................................................................... ix

List of Acronyms ............................................................................................................... xi

Chapter 1. Introduction ................................................................................................ 1

1.1. Eye Anatomy ........................................................................................................... 2

1.2. Diabetic Retinopathy ............................................................................................... 4

1.3. Glaucoma ................................................................................................................ 5

1.4. Ophthalmic Imaging Modalities ............................................................................... 5

1.4.1. Fluorescein Angiography ................................................................................ 6

1.4.2. Optical Coherence Tomography Angiography ................................................ 6

1.5. Contributions ........................................................................................................... 6

1.6. Outline of Thesis ..................................................................................................... 8

Chapter 2. Research Motivation ................................................................................. 9

2.1. Methods .................................................................................................................. 9

2.1.1. Visual Fields, Disc Photographs, Peripapillary Optical Coherence Tomography .................................................................................................................. 9

2.1.2. Optical Coherence Tomography Angiography Instrumentation .................... 10

2.1.3. Processing of OCT-A Images ....................................................................... 10

2.1.4. Image Acquisition and Quantification ............................................................ 11

2.2. Results .................................................................................................................. 13

2.3. Discussion ............................................................................................................. 15

2.4. Summary ............................................................................................................... 16

Chapter 3. OCT-A Image Strip-based Registration ................................................. 17

3.1. Methods ................................................................................................................ 19

3.1.1. Optical Coherence Tomography Instrumentation ......................................... 19

3.1.2. En face Angiogram Extraction ....................................................................... 19

3.2. Angiogram Registration ......................................................................................... 20

3.2.1. Microsaccade Free Strip Generation ............................................................ 20

3.2.2. Strip-based Affine Registration ..................................................................... 22

3.2.3. Strip-based Non-Rigid Registration ............................................................... 23

3.2.4. Validation ...................................................................................................... 24

3.3. Results .................................................................................................................. 25

3.4. Discussion ............................................................................................................. 29

3.5. Summary ............................................................................................................... 30

vii

Chapter 4. Automated Quantification ....................................................................... 32

4.1. Methods ................................................................................................................ 32

4.1.1. Inclusion Criteria ........................................................................................... 33

4.1.2. Optical Coherence Tomography Instrumentation ......................................... 33

4.1.3. Imaging Protocols ......................................................................................... 33

4.1.4. Processing of OCT-A Images ....................................................................... 33

4.1.5. Manual Tracing Methods ............................................................................... 34

4.1.6. Algorithm Training Methods .......................................................................... 34

4.1.7. Segmentation Performance Analysis ............................................................ 35

4.1.8. Clinical Outcome Measures .......................................................................... 35

4.2. Results .................................................................................................................. 36

4.2.1. Deep Neural Network Algorithm Performance .............................................. 36

4.2.2. Clinical Outcome Measures .......................................................................... 37

4.3. Discussion ............................................................................................................. 39

4.4. Summary ............................................................................................................... 42

Chapter 5. Future Work ............................................................................................. 43

5.1. 4D Registered OCTA Volumes ............................................................................. 43

5.2. Registration of Photoreceptor Images ................................................................... 43

5.3. Apply the DNN Framework to Averaged Images .................................................. 44

References ..................................................................................................................... 46

Appendix. Further Examples of Averaged OCT-A Images ............................... 54

viii

List of Tables

Table 1. Clinical Outcome Measures .............................................................................. 38

ix

List of Figures

Figure 1.1 Representative images of the Indian Ocean seen by someone with A) normal vision, B) diabetic retinopathy, and C) glaucoma. ......................... 1

Figure 1.2 Eye diagram [credit: National Eye Institute, National Institutes of Health] . 2

Figure 1.3 A portion of the human retina. Transverse histological retinal section was stained with toluidine blue (A), and a B-scan image was acquired using OCT (B) to illustrate the various retinal layers at the eccentricity located 3 mm superior to the optic disk. Colored dashed lines demarcate the retinal layers. Orange dashed lines indicate NFL; red dashed lines, RGC capillary network; yellow dashed lines, capillary network at IPL/sINL border; green dashed lines, capillary network at the dINL/OPL border. Scale bar: 50 μm. Image from [1]. ................................................. 3

Figure 1.4 Fundus photograph showing fluorescein imaging of the major arteries and veins in a human left eye. .......................................................................... 4

Figure 1.5 An example of OCTA images processed using the first version of the processing pipeline (left), and the current version of the pipeline (right). .. 7

Figure 2.1 Manual tracing techniques for quantifying radial peripapillary capillary (RPC) density. Radial peripapillary capillaries are seen in the speckle variance OCT-A image of a normal eye (Top). Manual tracing of RPCs (Center; red) were performed, the results of which were used to express the density of RPCs as a percentage of the total tissue area (Bottom). Note that large vessels were excluded from the tracing. Scale bar = 300 μm............................................................................................................ 12

Figure 2.2 Structural changes to radial peripapillary capillaries (RPCs) in unilateral glaucoma. The right optic disc (Top row, first image) demonstrates a myopic tilt however the automated Humphrey visual field test (Second row, first image) appears normal. Speckle variance OCT-A images of RPCs (Third row, first image) and the deep capillary plexus (Fourth row, first image) in the superotemporal peripapillary region are within the normal range. The left glaucomatous eye also demonstrates tilting (Top row, second image) but an inferior field defect is seen on visual field examination (Second row, second image). There is loss of RPCs in the superotemporal peripapillary region (Third row, second image) as seen on the speckle variance OCT-A image. The deeper capillary plexus at sites of RPC loss however appears normal and comparable in morphology to the fellow eye (Fourth row, second image). Projection artifacts from the large retinal vessels within the inner retina could be seen in the deeper capillary plexus images in both eyes (Fourth row, first and second image). Scale bar = 300μm. ................................................ 14

Figure 3.1 Overview of the strip-based registration algorithm for multiple serially acquired OCT-A images. Representative images are used to demonstrate the algorithm in Figure 3.2. ...................................................................... 20

Figure 3.2 Demonstration of the image stripping, coarse translation, affine registration and non-rigid registration steps of the proposed algorithm. The template image (green) and registered strip (magenta) are shown as composite images where white regions indicate where the two images

x

have the same intensities. The areas under the red, orange and yellow boxes are further explored in Figure 3.3. ................................................. 21

Figure 3.3 Comparison of three different strips registered to the same template image using (a) coarse translation, (b) affine registration, and (c) non-rigid registration. The template image (green) and registered strip (magenta) are shown as composite images where white regions indicate where the two images have the same intensities. .................................................... 22

Figure 3.4 Template image, mean, and median averaged images (all retinal layers, superficial and deep plexus) for Subject 3 OD, a healthy male subject, 29 years of age. ............................................................................................ 26

Figure 3.5 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age. .......................................................................................................... 27

Figure 3.6 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a subject with diabetic retinopathy. . 28

Figure 3.7 SSIM values for incremental averaged images for all eyes for the all retinal layers, superficial layers and deep layers. .................................... 29

Figure 4.1 An example image from both the prototype and commercial systems with their corresponding manual and automated segmentations. As some data within the commercial dataset contained an icon in the lower left corner a mask was applied and can be seen in the lower left corner of the automated segmentation result. .............................................................. 37

Figure 4.2 FAZ perimeter (yellow), maximum diameter (green) and minimum diameter (red) shown for example healthy and diabetic data from both systems using both manually and automated segmentations. As some data within the commercial dataset contained an icon in the lower left corner a mask was applied and can be seen in the lower left corner of the commercial automated segmentation results (white) and the lower left corner in the manual segmentation of the commercial diabetic image (black). ..................................................................................................... 39

Figure 4.3 Examples of low quality input data and the automated segmentation. Due to the low signal-to-noise ratio within the FAZ, some areas were erroneously segmented. Additionally, a horizontal motion artefact can be seen cutting through the FAZ which was also incorrectly segmented in areas. ....................................................................................................... 42

Figure 5.1 Two AO-OCT images of photoreceptors acquired from the same patient in a similar area. Scale bar is 50μm. ....................................................... 44

xi

List of Acronyms

CNR

DNN

DR

FA

Contrast to Noise Ratio

Deep Neural Networks

Diabetic Retinopathy

Fluorescein Angiography

FAZ

FWHM

ICC

ILM

INL

IOP

IPL

OCT

OCT-A

ONH

ONL

OPL

RPC

SIFT

SNR

SSADA

SSIM

VF

Foveal Avascular Zone

Full Width Half Maximum

Intraclass Correlation Coefficients

Inner Limiting Membrane

Inner Nuclear Layer

Intra-Ocular Pressure

Inner Plexiform Layer

Optical Coherence Tomography

Optical Coherence Tomography Angiography

Optic Nerve Head

Outer Nuclear Layer

Outer Plexiform Layer

Radial Peripapillary Capillaries

Scale Invariant Feature Transform

Signal to Noise Ratio

Split-spectrum amplitude-decorrelation angiography

Structural Similarity Index Measure

Visual Field

1

Chapter 1. Introduction

Vision is one of the five senses that many people take for granted every day.

Retinal diseases are one of the leading causes to affect the light sensitive tissue at the

back of the eye, affecting the ability to detect light. Two of the most common eye

diseases that cause blindness are diabetic retinopathy and glaucoma. Simulated

examples of the impacts on vision from these diseases are shown in Figure 1.1. Diabetic

Retinopathy (DR) causes a partial blurring or patchy loss of vision as shown in Figure

1.1B. Patients with glaucoma, on the other hand, may experience loss of peripheral

vision, called tunnel vision, as depicted in Figure 1.1C. As our population ages, the

number of people affected by these diseases is expected to rise significantly. The

National Eye Institute estimates that number of people who will have DR will nearly

double from 7.7 million to 14.6 million and the number of people with glaucoma will more

than double from 2.7 million to 6.3 million from 2010 to 2050 in the United States. This

increases the need for better diagnostic tools which enable ophthalmologists to detect

these diseases earlier and with better confidence.

Figure 1.1 Representative images of the Indian Ocean seen by someone with A) normal vision, B) diabetic retinopathy, and C) glaucoma.

Over the past decade, the development of visible and near-infrared retinal

imaging technology has grown rapidly. One of the dominant imaging modalities is

Fourier Domain Optical Coherence Tomography (FDOCT), which has revolutionized

clinical diagnostic ophthalmic imaging. FDOCT provides a detailed volumetric view of the

retina for clinicians to identify the structural hallmarks of diseases such as DR and

2

glaucoma. FDOCT images are used to assess the need for treatments (surgical,

intravitreal injection, laser, etc.) and afterwards to evaluate the results and monitor

changes.

In the remainder of this chapter, an overview of the eye anatomy, with a focus on

the retina is presented. This is followed by a detailed description of DR and glaucoma.

The imaging technologies that are specifically used by clinicians for visualizing the

retinal vasculature are described. The chapter concludes with an overview of the rest of

this thesis, and my contributions towards the development of tools for the betterment of

Optical Coherence Tomography Angiography (OCT-A) images.

1.1. Eye Anatomy

The eye is a complex organ that allows us to perceive and convert light to

electrical signals that the brain can interpret. Figure 1.2 presents a simple schematic of a

human eye.

Figure 1.2 Eye diagram [credit: National Eye Institute, National Institutes of Health]

3

Briefly, a collimated beam incident on the eye is focused by the cornea and lens

onto the retina. The eye is roughly 25mm in diameter, and the retina, located at the back

of the eye contains cell layers that detect light, perform some processing on the

information, and transmit electrical signals to the brain via the neurons of the optic nerve.

The rest of the report will focus mainly on the retina, a cross sectional diagram of which

is shown in Figure 1.3.

Figure 1.3 A portion of the human retina. Transverse histological retinal section was stained with toluidine blue (A), and a B-scan image was acquired using OCT (B) to illustrate the various retinal layers at the eccentricity located 3 mm superior to the optic disk. Colored dashed lines demarcate the retinal layers. Orange dashed lines indicate NFL; red dashed lines, RGC capillary network; yellow dashed lines, capillary network at IPL/sINL border; green dashed lines, capillary network at the dINL/OPL border. Scale bar: 50 μm. Image from [1].

There are two sources of blood supply to the human retina: the central retinal

artery (CRA) and the choroidal blood vessels. The inner retina is bounded by the nerve

fibre layer (NFL) anteriorly and the inner nuclear layer (INL) posteriorly, and it gets

nourished by blood that travels through the CRA from the optic nerve head (ONH). The

choroidal blood vessels feed the outer retina, particularly the photoreceptors.

The CRA follows the patterns as shown in Figure 1.4, where the vessels radiate

outward from the ONH and curve towards and around the fovea. These vessels supply

three layers of capillary networks which are the radial peripapillary capillaries (RPCs),

the inner and outer layer of capillaries. The RPCs are the most superficial layer of

capillaries lying in the inner part of NFL, and feed the superficial nerve fibres surrounding

the ONH. The inner capillaries lie in the GCLs under and parallel to the RPCs. The outer

4

capillary network runs from the inner plexiform layer (IPL) to the outer plexiform layer

(OPL) through the inner nuclear layer (INL) [8].

Figure 1.4 Fundus photograph showing fluorescein imaging of the major arteries and veins in a human left eye.

1.2. Diabetic Retinopathy

Diabetic retinopathy (DR) is the most prevalent retinal vascular disease

worldwide, affecting a third of people with diabetes [2]. It is a leading cause of adult

blindness, responsible for 15-17% of cases of blindness in the western world [3]. The

pathophysiology of DR is closely related to its deleterious effect on the inner retinal

microcirculation which includes altered vascular permeability and capillary bed closure

[4], [5]. Retinal ischemia secondary to capillary non-perfusion has been observed in the

early stages of diabetic retinopathy and has been correlated to disease severity and

progression [6]. Findings such as decreasing macular capillary density and enlargement

of the perifoveal zone [7] have been correlated with the severity of vision loss [8], [9].

Thus, early detection and reliable quantification of these microvascular changes may

play a role in predicting visual morbidity and improve the management of DR.

5

1.3. Glaucoma

Glaucoma is a leading cause of irreversible blindness worldwide[10] and the

second most common cause of blindness in the developed world [11]. The

pathophysiology of glaucoma is complex and characterized by the time-dependent loss

of retinal ganglion cells (RGCs) and their accompanying axons [12]. Indices that are

currently used to quantify and evaluate progression of glaucomatous optic neuropathy

include visual field testing, nerve fibre layer (NFL), optic nerve head, ganglion cell layer

with inner plexiform layer (GCIPL) and ganglion cell complex parameters analysis, and,

recently, measurement of lamina cribrosa thickness [13]. The RPCs represent a unique

capillary plexus within the inner aspect of the NFL. They are largely restricted to the

posterior pole of the human retina along specific retinal eccentricities surrounding the

optic nerve. Morphologically, this capillary network displays minimal inter-capillary

anastomosis and show a linear course in keeping with the NFL distribution. The

anatomical distribution and unique morphological characteristics help to distinguish the

RPCs from other capillary plexuses within the retinal microcirculation [12], [14]. The

nutritional demands of RGC axons are likely to be partially satisfied to a large extent

satisfied by radial peripapillary capillaries (RPCs) [15] and structural changes to RPCs

have been implicated in the pathogenesis of glaucoma [16]. Despite the evidence that

RPCs are critically related to RGC function [17]–[19], the morphological characteristics

of RPCs are not routinely used in clinical practice to evaluate glaucomatous progression.

This may be because RPCs are not reliably visualized with fluorescein angiography (FA)

[20] which is the mainstay imaging modality for clinically evaluating the retinal circulation.

1.4. Ophthalmic Imaging Modalities

The unique properties of the eye make it suitable for non-invasive optical

imaging. Imaging blood flow is very important because abnormal circulation is the

leading cause of irreversible blindness in diseases such as DR. Here we focus on

fluorescein angiography (FA) and Optical Coherence Tomography Angiography (OCT-

A).

6

1.4.1. Fluorescein Angiography

For the past 30 years, fluorescein angiography (FA) has been the gold standard

modality for assessing retinal vascular diseases [17] but vessel leakage and excessive

choroidal fluorescence affects the ability of FA to visualize the retinal microcirculation.

FA is an invasive procedure that requires venipuncture and the administration of

exogenous contrast agents. The injected dye can cause nausea, vomiting, skin

discolouration, pruiritis and in rare cases death and anaphylaxis [46]. Moreover, these

techniques only provide 2D information (en face views of the vasculature). Therefore,

there is a clinical demand for a non-invasive approach to provide visualization of the

microvasculature within the retina layers.

1.4.2. Optical Coherence Tomography Angiography

Optical Coherence Tomography Angiography (OCT-A) is an emerging imaging

modality with which the retinal circulation can be visualized by computing the

decorrelation signal on a pixel-by-pixel basis. Variants of OCT-A methods have been

described in recent review articles [21]–[24]. The speckle variance approach to OCT-A

has been evaluated against standard invasive techniques such as Fluorescein

Angiography (FA) [25], [26], in which only the superficial capillaries can be distinguished

due to excessive choroidal fluorescence[27], and ex vivo histological analyses [1], [26],

[28], [29].

1.5. Contributions

At the early stage of my graduate research, my work was mainly focused on data

acquisition, where I became familiar with the potential uses and limitations of OCT

Angiography. I was able to acquire and help analyze OCT-A data of the Optic Nerve

Head (ONH) looking closely at the RPCs which contributed to a published paper [30].

Through analyzing this data, I took the first version of the OCT-A processing pipeline

(which was on a single computer in the BORG lab at SFU and required the use of a

special version of the custom ‘OCTViewer’ acquisition software) and adapted it to work

on a computer cluster server so that any team member could process the data. Although

the first version of the code was useful for research, the inability to parallelize the

processing made it too slow to produce a clinical output in one day. Additional layers

7

were also added to the graph cut segmentation codes and additional filters were added

to emphasize the details of the fine microvasculature. This can be seen in Figure 1.5, in

which the capillaries have a higher contrast (qualitatively) with the updated processing

methods.

Figure 1.5 An example of OCTA images processed using the first version of the processing pipeline (left), and the current version of the pipeline (right).

By working on this study, two needs became apparent: 1) better quality en face

OCT-A angiograms, and 2) a more automated process for quantifying the data was

needed for accurate clinical diagnosis. Through working with an exchange PhD student,

a process for segmenting the retinal vasculature in our OCT-A data using Deep Neural

Networks (DNNs) was developed which led to two co-first authored papers [31],[32]. My

main first-author paper [33] demonstrated an automated method for registration and

averaging of serially acquired OCT-A images. The improved visualization of the

capillaries will hopefully enable more robust quantification and study of minute

changes in retinal microvasculature in the future.

8

Although my main contributions pertain to our OCT-A studies [26], [30]–[34], I

was also able to contribute to some of the adaptive optics [35],[36] anterior segment [37]

and morphological analysis [38],[39] work in a lesser capacity.

1.6. Outline of Thesis

The remainder of this thesis is organized as follows. In Chapter 2, motivation for

the research is presented in the form of a clinical study which utilizes OCT-A to image

the RPCs in focal glaucoma. Chapter 3 details an image processing pipeline for the

registration and averaging of serially acquired OCT-A images. Chapter 4 investigates the

ability of machine learning for automated analysis of OCT-A data. Lastly, the thesis ends

with a summary and future work.

9

Chapter 2. Research Motivation

Previous studies evaluated the morphological characteristics of the foveal [26],

perifoveal [28] and peripapillary capillary [29] networks using OCT-A [25] and showed

that the topological and quantitative characteristics of these networks, as seen on OCT-

A, are comparable to histologic representation. This chapter utilizes OCT-A to

quantitatively evaluate RPCs in glaucoma, glaucoma suspects, and normal eyes which

provides clinical motivation for the research presented in this thesis.

2.1. Methods

The study was designed as a prospective observational case series. All subject

recruitment and imaging took place at the Eye Care Centre at Vancouver General

Hospital. The study protocol including subject recruitment and imaging was approved

prospectively by the Research Ethics Boards at the University of British Columbia and

Vancouver General Hospital. The study was performed in accordance and adhered with

the tenets of the Declaration of Helsinki. Written informed consent was obtained from all

subjects.

2.1.1. Visual Fields, Disc Photographs, Peripapillary Optical Coherence Tomography

Visual fields were acquired using the Humphrey Field Analyzer II (Carl Zeiss

Meditec, Dublin, CA). Refractive error was corrected during testing. Stereoscopic photos

around the optic discs were obtained for each participant using a fundus camera (TRC-

50DX; Topcon, Japan) with 5.0-megapixel resolution. Peripapillary assessment of the

NFL was done using the standard peripapillary protocol using SD-OCT (Spectralis,

Heidelberg Engineering, Germany). All ancillary testing were acquired within six months

of OCT-A imaging.

10

2.1.2. Optical Coherence Tomography Angiography Instrumentation

Speckle variance OCT-A images and simultaneous, co-registered regular

structural OCT images were acquired from a GPU-accelerated OCT clinical prototype.

The details of the acquisition system have previously been published [25]. The OCT

system used a 1060nm swept source (Axsun Inc.) with 100 kHz A-scan rate and a full-

width half-maximum bandwidth of 61.5nm which corresponded to a coherence length of

~6μm in tissue. The size of the focal waist on the retina was estimated using the

Gullstrand-LeGrand model of the human eye to be ω0 ≈ 7.3 μm (calculated using

Gaussian optics) corresponding to a lateral FWHM of ~8.6 μm. For the angiogram, the

speckle variance calculation [40] was used

2

1 1

1 1( )

N N

jk ijk ijk

i i

sv I IN N

,

(1)

where i, j, and k are the indices of the frame, width, and axial position of the B-scan

respectively, I is the intensity at the index and N is the number of repeat acquisitions per

BM-scan (N=3). Processing of the OCT intensity image data and en face visualization of

the retinal microvasculature was performed in real time using our open source code for

alignment and quality control purposes [41], [42]. The scan area was sampled in a

300x300(x3) grid with a ~2x2mm field of view in 3.15 seconds. Scan dimensions were

calibrated based on the eye length of each participant, measured using the IOL Master

500 (Carl Zeiss Meditec Inc., Dublin, California, USA).

2.1.3. Processing of OCT-A Images

Post-processing of the raw intensity data was performed to segment the retinal

layers and extract optimal quality images of the retinal microvasculature. Coarse axial

motion artifact was corrected using cross-correlation between adjacent frames. Sub-

pixel registration was performed on each set of corresponding B-scans before creating

the speckle variance B-mode scan. Before layer segmentation, three-dimensional

bounded variance smoothing was applied to the motion corrected intensity images in

order to reduce the effect of speckle while preserving and enhancing edges. The inner

limiting membrane, posterior boundary of the NFL, inner nuclear layer and outer nuclear

11

layer were segmented automatically in 3D using a graph-cut algorithm [43]. The

automated segmentation was examined and corrected by a trained grader using Amira

(version 5.1; Visage Imaging, San Diego, CA, USA). The OCT-A image within each layer

was summed in the axial direction to produce a projected en face image. En face images

were notch filtered and contrast-adjusted using adaptive histogram equalization. In an

effort to eliminate the bias of large blood vessels from the NFL thickness measurement,

the images were cropped to an area of ~1x1mm to remove the majority of the large

blood vessels present. Furthermore, prior to performing capillary density comparisons

between glaucoma, glaucoma suspect and normal control subjects the images were

again cropped to an area of 636.5x636.5μm.

2.1.4. Image Acquisition and Quantification

Group A consisted of subjects with unilateral glaucoma, Group B consisted of

glaucoma suspects, and Group C consisted of healthy subjects. The average age of

subjects (range and median) of Groups A, B and C were 58.00±18.60 (26-72: 66) years,

44.67±23.50 years (21-68: 45) and 41.13±13.51 years (27-60: 36), respectively

(P=0.25). The male: female ratio of subjects in Group A, B and C were 3:2, 3:0 and 4:3,

respectively. The average intraocular pressure of subjects (range and median) of the

glaucoma eyes and fellow eyes in Group A and Group B were 11.46 ± 4.21 mmHg (5-

15: 12), 13.74 ± 3.21 mmHg (10-17: 14) and 16.17 ± 2.04 mmHg (14-19: 16),

respectively.

Our previously published manual tracing technique was used to quantify RPC

density as shown in Figure 2.14 All manual tracings were performed by ZM in a non-

blinded fashion. Manual tracing was performed using the GNU Image Manipulation

Program Version 2.8.14. Care was taken to trace RPCs, and all large vessels

originating from the disc were segmented separately and excluded. Capillary density

was measured in the segmented image using MATLAB. The proportion of the image

occupied by retinal vessels was expressed as a percentage, and the unit of

measurement was calculated as the percentage retinal area occupied by capillary

plexus. As speckle variance images are derived from the structural intensity scans,

quantitative structural information such as the NFL thickness can be extracted from the

exact same location as capillary density. The NFL thickness was measured and

averaged across the cropped volumetric ~1x1mm scan of the region of interest. Where

12

reported, the density of the RPCs was calculated as the fraction of pixels identified as

belonging to a vessel versus the total number of pixels in an image. To determine the

reproducibility of these measurements, three images from Group A, B and C were

manually traced and quantified on two separate occasions in a blinded fashion, each at

least 3 months apart, by the same rater (ZM). To facilitate the qualitative comparisons of

the deep capillary plexus, careful manual segmentation of the retinal layers was

performed to minimize projection artifacts. In addition, manually scrolling through the

entire OCT volume helped to identify and exclude any residual projection artifacts from

the overlying large superficial vessels in the qualitative comparison.

Figure 2.1 Manual tracing techniques for quantifying radial peripapillary capillary (RPC) density. Radial peripapillary capillaries are seen in the speckle variance OCT-A image of a normal eye (Top). Manual tracing of RPCs (Center; red) were performed, the results of which were used to express the density of RPCs as a percentage of the total tissue area (Bottom). Note that large vessels were excluded from the tracing. Scale bar = 300 μm.

13

2.2. Results

In the normal control group, the RPCs followed a very similar trajectory to the

RGC axons in the NFL and demonstrated a linear course with minimal anastomoses.

Decreased density of RPCs was observed within regions of optic disc neural rim loss in

glaucomatous eyes. In glaucomatous eyes, RPCs maintained a linear trajectory,

however a patchy or diffuse loss of RPCs was observed within regions of NFL thinning.

The density and morphologic characteristics of deeper capillary networks, beyond the

outer margins of the NFL, at sites of RPC loss appeared normal in glaucomatous eyes

(Figure 2.2).

14

Figure 2.2 Structural changes to radial peripapillary capillaries (RPCs) in unilateral glaucoma. The right optic disc (Top row, first image) demonstrates a myopic tilt however the automated Humphrey visual field test (Second row, first image) appears normal. Speckle variance OCT-A images of RPCs (Third row, first image) and the deep capillary plexus (Fourth row, first image) in the superotemporal peripapillary region are within the normal range. The left glaucomatous eye also demonstrates tilting (Top row, second image) but an inferior field defect is seen on visual field examination (Second row, second image). There is loss of RPCs in the superotemporal peripapillary region (Third row, second image) as seen on the speckle variance OCT-A image. The deeper capillary plexus at sites of RPC loss however appears normal and comparable in morphology to the fellow eye (Fourth row, second image). Projection artifacts from the large retinal vessels within the inner retina could be seen in the deeper capillary plexus images in both eyes (Fourth row, first and second image). Scale bar = 300μm.

15

2.3. Discussion

Radial peripapillary capillaries comprise a unique vascular plexus that is

predominantly found in the posterior pole of primates with a typical macula. The

metabolic demands of peripapillary RGC axons are likely to be mainly nourished by the

RPCs. There is also evidence to demonstrate an association between RPC loss and

NFL changes in chronic glaucoma. Although clinical imaging of RPCs may be a

potentially useful way for evaluating and monitoring RGC axonal disease it is not

routinely used in clinical practice due to the difficulties associated with visualizing this

circulation using FA. Our previous study showed that quantitative analysis of retinal

capillary detail could only be performed in 30% of FA images acquired from normal

subjects with clear ocular media. The major limiting factor that precludes clear

visualization of retinal capillaries on FA is fluorescence from the choroidal circulation.

Our recent studies have quantified the morphological characteristics of retinal capillary

networks as seen OCT-A and have shown that it is comparable to histological

representation thus suggesting that OCT-A techniques may be useful for evaluating the

structural characteristics of retinal capillary networks. Optical coherence tomography

angiography overcomes some of the limitations of FA as it is a label-free technique that

permits non-invasive, depth-resolved evaluation of retinal capillary networks [22]. As

shown in Figure 2.2, RPC loss was identified in the NFL in glaucomatous eyes while the

deeper capillary networks appeared structurally normal.

Manual tracing techniques were used to calculate RPC density and in order for

this technique to have broad clinical utility an automated method for determining RPC

density will be required. Skeletonization algorithms and binary image analysis

techniques may potentially overcome this limitation. All manual tracings were performed

by ZM in a non-blinded fashion to minimize grader bias, all tracings were collectively

reviewed and approved by ZM, MH and CB prior to analysis. Projection artifacts could

have affected our qualitative comparison of the deep capillary networks between the

study groups. As outlined in the methods section, a number of manual steps were taken

to minimize the potential deleterious effects of projection artifacts from overlying vessels.

This topic is receiving increased attention in the OCTA literature, automated methods for

removing projection artifacts are still in the development phase.

16

2.4. Summary

In this chapter, OCT-A was used to help answer a clinical question proving the

clinical utility of this technology; however, there were also several limitations noted which

could be improved upon – specifically the low image quality and use of manual

quantification.

In the next chapter, a pipeline for improving the visible definition of retinal

microvasculature in OCT-A by motion correction, registration, and averaging of

sequentially acquired images will be described. Detailed, high quality OCT-A images are

needed for clinical studies such as comparisons of OCT-A with histology and fundus

photography fluorescein angiography (FA), and studying the shunting of vessels in a

focal area, such as the inner ring of vessels in the foveal avascular zone (FAZ), or in

glaucomatous focal defects as shown in this chapter.

17

Chapter 3. OCT-A Image Strip-based Registration

Non-invasive, in vivo visualization of the retinal microvasculature using OCT-A

can be instrumental in studying the onset and development of retinal vascular diseases.

For example, OCT-A has enabled the visualization of the deep plexus layer and

furthered the understanding of diseases such as paracentral acute middle maculopathy

[44]–[46] and diabetic retinopathy [47], [48]. Quantitative measurements, such as

capillary density, can be used to stratify the risk of disease progression, visual loss, and

also for monitoring the course of disease [9], [49]. As mentioned in the previous chapter,

due to projection artifact and poor contrast it is often difficult to trace individual vessels in

this layer when only one en face image is visualized. An additional challenge to this end

is the small dimension and pulsatile flow of the retinal capillaries, making them less

consistently visible and difficult to distinguish from the speckle noise relative to larger

vessels. This limits the detection sensitivity for changes in the retinal microvascular

circulation due to diseases, aging, or treatment. Methods for reliable visualization of the

microvasculature in the OCT-A images are required for studies conducting longitudinal

and cross-sectional quantitative analysis. Detailed, high quality OCT-A images are

needed for clinical studies such as comparisons of OCT-A with histology and fundus

photography fluorescein angiography (FA), and studying the shunting of vessels in a

focal area, such as the inner ring of vessels in the foveal avascular zone (FAZ), or in

glaucomatous focal defects [30].

Serially acquiring and averaging multiple OCT-A images can be an effective

solution for confirming the presence or absence of capillaries as the discontinuous

appearance of the capillary vessels is beyond improvement simply by just applying

image filtering [23], [50]. A crucial step in the serial acquisition approach is the

registration of multiple OCT-A images, the difficulty of which is compounded by the fact

that an OCT-A image is acquired over multiple seconds and thus particularly susceptible

to motion artifacts. The registration of sequential B-scans [51] can aid in attenuating

small motion artifacts, but not the larger motion artifacts associated with imaging

subjects with pathologies. OCT-A with eye-tracking has been implemented in

commercial retinal imaging systems, although this increases hardware cost and

18

complexity on which the sensitivity and reliability of motion detection also depend.

Previous works on post hoc motion artifact removal by en face summed volume

projection (SVP) have been reported in the Literature, see for example [23], [52], [53]. In

Hendargo et al. [52], two to three sets of orthogonal (x-fast and y-fast) volumes were

acquired and divided into motion-free strips. The visualization and contrast of the

vessels were improved by multi-resolution Gabor filtering, and the strips were registered

one-by-one, first globally by x- and y- translation that maximized the correlation in the

overlapping region, and locally by B-spline free-form deformation in the overlapping

region. Zang et al. [53] did not acquire orthogonal data sets, but instead serially acquired

two OCT-A volumes in the same scan orientation that were divided into parallel motion-

free strips. The strips were first registered by x- and y-translation and rotation that

minimized the squared difference of ‘large vessels’, which were defined in the paper as

pixels with decorrelation value greater than 1.3 times the mean value. This was followed

by B-spline free-form deformation on ‘small vessels’, defined as pixels with decorrelation

value less than 1.3 times and greater than 0.6 times the mean value. Both groups

presented mosaicking of OCT-A images into widefield views, which has been reported in

other works as well [51], [54].

In this Chapter, averaging of up to 10 serially acquired OCT-A images with

parallel strip-wise microsaccadic noise removal and localized nonrigid registration is

presented. Unlike the previous two methods [52], [53] which concentrated on motion

artifact removal and widefield imaging, our purpose was to improve the contrast and

signal to background of the capillaries in focal regions. The details of our methodology

are presented below. In brief, the serially acquired OCT-A images were divided into

microsaccade-free strips. The target strips were first aligned to a template image by x-

and y-translation based on maximum cross-correlation, followed by affine registration

using Scale Invariant Feature Transform (SIFT) [55], a feature extraction method robust

to scaling, orientation changes, illumination changes, and affine distortions.

The image warping and local distortion due to slower eye movements are less

obvious and more difficult to model than the strong stripe artifacts from microssacadic

motion. Instead of free-form deformation [52], [53], our approach optimized the intensity

value at each pixel location as the average of the values from each overlapping strip

determined by translation and rotation of a windowed region in each strip. Thus pixel-

19

wise correspondence across multiple OCT-A images was found by local neighborhood

matching.

The remainder of this Chapter is organized as follows. The Methods section

describes our processing algorithm for OCT-A image averaging in detail, as well as

quantitative metrics to evaluate the improvements to the image quality. The algorithm

was tested on OCT-A images of six healthy volunteers, with the vessel visibility

improvement qualitatively demonstrated in all, superficial, and deep plexus layers.

Quantitatively, the algorithm performance was evaluated by contrast to noise ratio (CNR)

and signal to noise ratio (SNR) with the background speckle noise information from the

foveal avascular zone (FAZ). The Chapter ends with a discussion of the image quality

improvements using our method of averaging serially acquired OCT-A images.

3.1. Methods

All subject recruitment and imaging took place at the Eye Care Centre of

Vancouver General Hospital. The project protocol was approved by the Research Ethics

Boards at the University of British Columbia, Simon Fraser University, and Vancouver

General Hospital, and performed in accordance with the tenets of the Declaration of

Helsinki. Written informed consent was obtained from all subjects.

3.1.1. Optical Coherence Tomography Instrumentation

The OCT-A images were acquired from a GPU-accelerated OCT clinical

prototype, previously described in Section 2.1.2. Ten serially acquired volumes centered

at the foveal avascular zone (FAZ) were obtained per eye in ~32s. During this image

acquisition period, patients were asked to maintain their gaze on a particular target, and

encouraged to blink as necessary in order to prevent drying of the cornea. The

automated parsing of the image data strips (Section 3.2.1) eliminated issues of motion

artifact and partial volumes.

3.1.2. En face Angiogram Extraction

Post-processing of the raw intensity data was performed to extract optimal quality

images of the retinal microvasculature according to the procedure outlined in Section

20

2.1.3. Projection artifacts in the deep layer angiogram were attenuated using a modified

slab-subtraction algorithm [56]. In Equation (2, 𝑃𝑅𝐷𝑒𝑒𝑝 is the projection resolved en face

angiogram of the deep layer, where Norm⟨… ⟩ represents the normalization process,

𝑁𝐷𝑒𝑒𝑝 is the number of pixels in the deep layer, 𝑁𝐴𝑙𝑙 𝐿𝑎𝑦𝑒𝑟𝑠 is the number of pixels in all

the retinal layers, 𝑁𝑆𝑢𝑝𝑒𝑟𝑓𝑖𝑐𝑖𝑎𝑙 is the number of pixels in the superficial layer, and 𝑠𝑣 is the

angiogram.

Deep Deep All Layers SuperficialDeep All Layers Superficial

1 1 1Norm * Norm NormPR sv sv sv

N N N

. (2)

3.2. Angiogram Registration

The algorithm overview is shown in Figure 3.1. The ten serially acquired en face

images of all retinal layers were divided into microsaccade free strips, which were then

registered to a template image, first using rigid registration for the course alignment,

followed by non-rigid registration for finer features. Transforms applied to the en face

image of the full retinal thickness were then applied to both the superficial and deep

layer angiograms.

Figure 3.1 Overview of the strip-based registration algorithm for multiple serially acquired OCT-A images. Representative images are used to demonstrate the algorithm in Figure 3.2.

3.2.1. Microsaccade Free Strip Generation

For each eye, a microsaccade-free image from the ten en face images was

chosen as the template image. In the case that all images contained microsaccadic

motion artifacts, a template was generated by stitching together microsaccade-free strips

using the registration methods discussed below.

21

After the template image was chosen / generated, the remaining images were

divided into strips between positions in the image corresponding to where the patient

fixation was lost, which appeared as vertical white stripes in the en face image, as

shown in Figure 3.2. Strips less than 40 pixels wide often contained large drift artifact

and were therefore discarded. If multiple microsaccade free images existed per eye, the

first was selected as the template and the rest were divided into three equal-sized strips

for registration. Each strip was zero-padded to match the size of the template image and

coarsely aligned to the template by x- and y-translation using maximum cross

correlation.

Figure 3.2 Demonstration of the image stripping, coarse translation, affine registration and non-rigid registration steps of the proposed algorithm. The template image (green) and registered strip (magenta) are shown as composite images where white regions indicate where the two images have the same intensities. The areas under the red, orange and yellow boxes are further explored in Figure 3.3.

22

Figure 3.3 Comparison of three different strips registered to the same template image using (a) coarse translation, (b) affine registration, and (c) non-rigid registration. The template image (green) and registered strip (magenta) are shown as composite images where white regions indicate where the two images have the same intensities.

3.2.2. Strip-based Affine Registration

Scale Invariant Feature Transform (SIFT) keypoints were automatically extracted

from both the template image and each strip to be registered [55]. Briefly, keypoints are

the locations of local scale-space extrema in the difference-of-Gaussian function

convolved with the image. Further refinement to the keypoints can be made by assigning

each keypoint an orientation to achieve invariance to image rotation. Finally, a local

image descriptor is assigned to each keypoint using the image location, scale and

orientation as found above. Readers are encouraged to refer to [55] for a more detailed

description of the SIFT algorithm.

As the SIFT feature descriptor is invariant to uniform scaling and orientation, it is

ideal for identifying matching keypoints in noisy or speckled images such as OCT-A

angiograms. The calculation of Euclidean distances in MATLAB is computationally

expensive, and therefore matching keypoints between the template and strip were

identified as the closest corresponding keypoints by a small angle approximation to the

Euclidean distance. Keypoints were considered matching if the ratio of the vector angles

from the nearest to the second nearest match was less than a threshold value of 0.75.

23

As the image had been coarsely aligned in the previous step, a second check was

included to ensure the matched keypoints were no more than 40 pixels distant in the x or

y direction.

All strips that had a minimum of 4 matched keypoints were then transformed

using an affine transform estimated using the matching keypoints as inputs to the

EstimateGeometricTransform function in MATLAB. This function iteratively compares an

affine transformation using three randomly selected keypoints, where the transformation

with the smaller distance metric calculated using the M-estimator SAmple Consensus

(MSAC) algorithm is used as the transformation matrix for the next comparison. The

maximum number of random trials for finding the inliers was set to 5000 for improved

robustness.

3.2.3. Strip-based Non-Rigid Registration

The vertical white lines in the target image in Figure 3.2 mark the image discontinuities

due to microsaccades accounted for by the strip-based affine registration; however,

localized mismatch still remains in the aligned images after this affine registration step.

The next step in our algorithm is to compensate for the smoother tremor and drift

motions represented by image warping and distortion, by using non-rigid registration.

Prior to non-rigid registration, a 2x2 averaging filter was applied to both the template and

the aligned strip to smooth any fine speckle that may affect the non-rigid registration.

The template and aligned strip were both then zero padded by 15 pixels. For each pixel

in the strip the normalized cross-correlation[57] was calculated, defined by,

,,

2 2

,, ,

[ ( , ) ][ ( , ) ]( , )

[ ( , ) ] [ ( , ) ]

s tx y

norm

s tx y x y

f x y f m x s y t mxcorr s t

f x y f m x s y t m

,

(3)

where f(x,y) is the 29x29 pixel matrix field centered on the (x,y) pixel of the template

image, 𝑓𝑠,𝑡̅̅ ̅̅ is the mean of the image in the region under the mask, m is the 15x15 pixel

mask matrix centered on the pixel of the strip and �̅� is the mean of the mask. This was

also done for -15,-10,-5, 5, 10, and 15° rotated field matrices. The pixel located at the

index of the maximum normalized cross correlation was then used as the registered

pixel for the strip. Figure 3.2 shows a pictorial schematic of the registration pipeline

24

described in this section. A smaller field of view demonstration of the coarse, affine and

non-rigid registration steps is shown in Figure 3.3. The stack of registered strips could

then either be combined by taking the mean or median to generate a higher quality

image.

3.2.4. Validation

The performance of the algorithm was evaluated with qualitative observation and

quantitative measures of the contrast to noise ratio (CNR), signal to noise ratio (SNR),

and structural similarity index (SSIM).

The CNR [58], [59] is defined as

2 210log r b

r b

CNR

,

(4)

where r and σr2 are the mean and variance of the whole image b and σb

2 are the mean

and variance of the background noise region. The background noise region was

selected to be the largest rectangle that would fit within the FAZ. As this is an area of

non-perfusion, any signal located here in a healthy eye can be considered noise. Vessel

segmentation to delineate the pure signal34 was not used here, as the quality metric

was only used for intra-volume comparison to measure the trends.

The SNR [58], [59] is defined as

2

2

max( )10log lin

lin

XSNR

,

(5)

where Xlin is the matrix of pixel values in the angiogram on a linear intensity scale and

σ2lin is the noise variance on a linear intensity scale. The background noise region

selected was the same used in the CNR calculations.

The SSIM [60] is a quality metric used to measure the perceived relative quality

of a digital image, and is defined by

25

1 2

2 2 2 2

1 2

(2 )(2 )( , )

( )( )

x y xy

x y x y

c cSSIM x y

c c

,

(6)

where x is the image to be compared, y is the final averaged image, and , σ2 and σ are

the average, variance and covariance respectively. The terms c1 and c2 are small

constants << 1 added to avoid instability when 2x + 2

y or σ2x + σ2

y are equal to zero.

3.3. Results

A total of 10 eyes from 6 healthy volunteers (4 male, 2 female) aged 36.8 ± 9.3

years were acquired according to the imaging protocol. A comparison of the template

image and the final averaged OCT-A images for all retinal layers, as well as the

superficial and deep vascular layers is shown in Figure 3.4 and Figure 3.5. In the

template images, the vessels near the FAZ are relatively clear; however, it becomes

harder to differentiate the vessels further towards the periphery. In contrast, the vessels

in the averaged images are clearly seen throughout. Improvement in vessel visibility is

particularly marked in the deep layer, where the OCT signal strength is weaker.

Qualitatively the median images appear sharper than the mean images as the median

averaging acts as a speckle reducing filter. However, the mean images appear more

smooth than the corresponding median image.

26

Figure 3.4 Template image, mean, and median averaged images (all retinal layers, superficial and deep plexus) for Subject 3 OD, a healthy male subject, 29 years of age.

27

Figure 3.5 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

28

Figure 3.6 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a subject with diabetic retinopathy.

For quantitative comparisons of the template and final averaged images, the

average CNR and SNR of the images was calculated. The average CNR of the

angiograms with all retinal layers increased from 0.52 ± 0.22 dB using the template

images to 0.77 ± 0.25 dB with the mean images, and 0.75 ± 0.24 dB with the median

images. Additionally, the average SNR of the angiograms with all retinal layers

increased from 19.58 ± 4.04 dB using the template images to 25.05 ± 4.73 dB with the

mean images, and 25.02 ± 4.89 dB with the median images. The mean improvement of

both the CNR and SNR was statistically significant (p<0.01) using a paired t-test.

To evaluate the change in perceptual quality per strip, the SSIM was calculated

on each incremental averaged image of the template and registered strips. Although 10

volumes were acquired per eye, the number of microsaccades and strips less than 40

29

pixels in the corresponding en face images was different for all eyes and therefore the

number of strips used to generate the averaged images was not necessarily equal. The

mean number of strips per eye was 21 ± 7 strips. As seen in Figure 3.7, the SSIM values

show a rapid increase as the first few strips are registered and applied to the template

image, and then the rate of improvement slows with additional registered strips. This

trend was observed in both the mean and median averaged images.

Figure 3.7 SSIM values for incremental averaged images for all eyes for the all retinal layers, superficial layers and deep layers.

3.4. Discussion

The major findings in this paper are as follows: averaging multiple registered

sequentially acquired OCT-A images (1) qualitatively enhances the visualization of the

retinal microvasculature networks, (2) increases the SNR and CNR of the angiograms,

and (3) increases the perceptual visual quality when using SSIM as a metric.

After averaging multiple en face images, the vessels of the deeper capillary

plexus are more readily identified, making quantification more reliable and thereby

facilitating investigation of its role in the pathophysiology of retinal vascular disease.

Although minimal projection artifact can still be seen in Figure 3.4 corresponding to the

30

larger superficial vessels, the overall qualitative condition of the en face images is

improved.

The SNR and CNR both increased significantly by averaging the individual strips.

Although both the SNR and CNR of the mean images are larger than those of the

median images, there is no significant difference between the mean and median and

therefore no recommendation of an averaging method can be made based on these

metrics.

The SSIM is a full reference metric where the final averaged image was taken to

be the perfect quality reference image. As shown in Figure 3.7, the SSIM increases with

each additional registered strip that is averaged to the template image. Note that each of

the 10 volumetric sets for each subject was divided differently into strips (based on the

motion), therefore a different number of strips was used for each averaged reference

image. As expected when averaging images, the first few strips applied to the template

affected the SSIM the most whereas the later strips provided only modest improvement

to the SSIM. The deep plexus showed the greatest increase overall. By increasing the

visibility of individual vessels, this technique has the potential to improve automated

segmentation results thereby improving our ability to quantify capillary density in normal

and diseased states.

Although the ability to enhance the visualization of the retinal plexuses through

averaging multiple sequentially acquired OCT-A images was demonstrated, there are

several limitations of this work which should be acknowledged. This study assessed only

relatively young subjects with clear ocular media and good fixation ability. The presence

of media opacities in older subjects may limit the amount of capillary information that can

be attained from images. Although the algorithm attenuates non-microsaccade motion in

the registered strips, the template may contain distortions and image warping which is

not accounted for here.

3.5. Summary

In this chapter, a pipeline for the registration of sequentially acquired OCT-A

images was presented to increase the quality of visualization. This improves one of the

31

main limitations in the previous chapter, and in the next chapter we will discuss how the

burned of manual segmentation was overcome using machine learning.

32

Chapter 4. Automated Quantification

Early efforts aimed at OCT-A perifoveal capillary density quantification utilized

manual vessel tracing methods [26], but the process can be labour intensive and subject

to intra- and inter-observer variability [61]. Fully automated techniques are being

explored, but face challenges such as variable intra- and inter-image signal to noise

ratios, projection artefacts from outer layer vasculature on to deeper layers, and motion

artefacts [31], [62]. OCT-A signal intensity thresholding has been the foundation of most

automated segmentation efforts thus far and progress has been made in applying

additional filters for more accurate results [49], [63]. Improvements in automated

segmentation and quantification of OCT-A images of the retinal vasculature may aid in

its wider spread adoption and potential application

Our group has demonstrated a novel automated deep learning method to

segment and quantify retinal images from a prototype OCT-A machine using Deep

Neural Networks (DNN) [31]. The application of the algorithm has been expanded to

include OCT-A images from a commercial system, the RTVue XR Avanti (Optovue, Inc)

[64]. In this Chapter, the use of the deep learning method for the quantification of Foveal

Avascular Zone (FAZ) parameters and perifoveal capillary density of prototype and

commercial OCT-A platforms in both healthy and diabetic eyes is evaluated.

4.1. Methods

The protocol for this study was approved by the human research ethics

committees of Simon Fraser University, the University of British Columbia and the North

Shore Long Island Jewish Health System and conducted in compliance with the

Declaration of Helsinki. Written informed consent was obtained from all subjects. Patient

imaging using the commercial device was performed at Vitreous Retina Macula

Consultants of New York from February 13, 2015 to April 25, 2016. Patient imaging

using the prototype device was performed at the Eye Care Centre in Vancouver, British

Columbia from July 14, 2014 to October 13, 2016. Data analysis for this study was

performed from February 25, 2016, to November 11, 2016. The subjects underwent a

33

standard ophthalmic examination and their level of retinopathy was determined by the

treating physician using the Early Treatment of Diabetic Retinopathy Study (ETDRS) [65]

staging.

4.1.1. Inclusion Criteria

Subjects classified as diabetic were diagnosed with diabetic retinopathy

according to the ETDRS criteria by an experienced retina specialist. Subjects that

comprised the control group showed no evidence of retinal or ocular pathology on

examination. All subjects were screened for clear ocular media, ability to fixate and were

able to provide informed consent prior to imaging.

4.1.2. Optical Coherence Tomography Instrumentation

Two OCT-A systems were used in this study: one prototype and one

commercially available machine. The clinical prototype OCT-A system used in this study

was previously described in Section 2.1.2. The commercial OCT-A system used in this

report was the RTVue XR Avanti (Optovue, Inc) which is an 840 nm spectral domain

system with an A-scan rate of 70 kHz. The XR Avanti has a reported axial resolution of

5μm and a transverse resolution of 15μm [66].

4.1.3. Imaging Protocols

Standard imaging procedures differed among the two OCT-A systems used. For

OCT-A using the prototype system, the protocol described in Section 3.1.2 was used.

For the RTVue-XR Avanti system, images were scanned over 3x3mm regions centered

on the FAZ with a scan pattern of 2 repeated B-scans at 304 raster positions, with each

B scan consisting of 304 A-scans. Two volumetric scans were acquired in this fashion:

one horizontally scanned and the other vertically for a total acquisition time of ~6.25s.

4.1.4. Processing of OCT-A Images

The commercial images were processed with the system’s built-in image

processing software, AngioVue. The split-spectrum amplitude-decorrelation angiography

(SSADA) method was used for extracting the OCT-A information. The algorithm split the

34

spectrum into 11 sub-spectra and detected blood flow by calculating the signal

amplitude-decorrelation between two consecutive B-scans of the same location. Both

horizontally acquired and vertically acquired images were registered and averaged.

4.1.5. Manual Tracing Methods

Two trained raters segmented OCT-A images using a Wacom Intuos 4 tablet

(SR), a Samsung ATIV Smart PC Pro 700T tablet (FC) and GNU Image Manipulation

Program (GIMP). Rater A (FC) segmented all prototype OCT-A images and 26 of the

commercial OCT-A images while rater B (SR) segmented the other 18 commercial OCT-

A images. The segmentations were reviewed and accepted by two other trained raters

(MH and ZM).

4.1.6. Algorithm Training Methods

The automated segmentation of the blood vessels in the OCT-A images was

performed by classifying each pixel into vessel or non-vessel class using deep

convolutional neural networks. A detailed description of the DNN architecture has been

previously published19. Briefly, original OCT-A en face images and the corresponding

manual segmentations were used as inputs to train the deep neural network. An equal

number of vessel and non-vessel pixels were extracted from each image to ensure a

balanced training set. The trained network then segmented the test datasets by

assigning a grayscale value, with higher values representing higher confidence of the

pixel being a vessel. The prototype and commercial devices were trained separately due

to inherent differences in the original images such as the image size. Both datasets were

trained on a mixed training set comprised of both healthy and diabetic images. The first

half of the dataset was used to train the deep neural network, which was then used to

segment the second half of the dataset. The process was repeated with the datasets

reversed. As the commercial training dataset comprised of two separate raters, care was

taken to divide the training sets equally between rater segmentations. Due to an icon in

the bottom left corner of some data in the commercial dataset, a mask was applied over

the area and it was disregarded in further analysis.

35

4.1.7. Segmentation Performance Analysis

The segmentation performance was evaluated by pixel-wise comparison of the

manually segmented images and the thresholded binary output of the deep neural

network using Otsu’s method for threshold selection. The accuracy, sensitivity, and

specificity were calculated for each pixel in the dataset and presented as a mean

average. For each dataset the number of true positive (TP), true negative (TN), false

positive (FP), and false negative (FN) pixels were used to calculate the accuracy

((TP+TN)/(TP+FP+FN+TN)), sensitivity ((TP/(TP+FN)), and specificity (TN/(TN+FP)).

4.1.8. Clinical Outcome Measures

Four FAZ morphometric parameters (area, maximum and minimum diameter,

and eccentricity) as well as perifoveal capillary density were calculated from the

automated segmentation results. The foveal vascular zone was found as the largest

connected non-vessel area. The centroid for this area was then used to determine the

maximum and minimum diameter. Eccentricity was calculated as 𝑒 = √1 −𝑏2

𝑎2 where 𝑏

is the minimum radius and 𝑎 is the maximum radius of the ellipse made by the maximum

and minimum diameter. Before capillary density measurements were calculated for the

automated segmentations a gamma correction filter was applied to ensure vessel

connectivity after binarization. Additionally, all erroneously segmented pixels within the

FAZ area were set to a non-vessel classification. Perifoveal capillary density was then

calculated as the proportion of measured area occupied by pixels which were classified

by the algorithm as a vessel.

Paired t-tests and Intraclass Correlation Coefficients (ICC) were used to compare

the means and agreement between segmentation methods, respectively, of the four FAZ

morphometric parameters and perifoveal capillary density. A Student’s t-test assuming

heteroscedasticity was used to compare automatically segmented eyes with and without

DR. Results for the prototype and commercial OCT-A systems were assessed

separately.

36

4.2. Results

A total of 71 eyes from 42 subjects were imaged as per the study protocol. 12

healthy subjects (21 eyes) and 5 diabetic subjects (7 eyes) were imaged with the

prototype OCT-A system, while 18 healthy subjects (31 eyes) and 7 diabetic subjects

(12 eyes) were imaged using the commercial RTVue XR Avanti (Optovue Inc., Fremont,

CA) system. The mean age of the prototype subjects was 37.3 ± 12.5 years, consisting

of 7 females and 10 males. The mean age of the commercial subjects was 37.9 ± 12.4

years, consisting of 9 females and 16 males. Of the diabetic eyes imaged with the

prototype system, the number of eyes with non-proliferative diabetic retinopathy (NPDR)

=3, Mild NPDR = 1, Severe NPDR = 2 according to the ETDRS Grading scheme, with 3

eyes exhibiting macular edema and 4 which had previous treatment for DR (laser and/or

IVI). Similarly, for the commercial system, the number of eyes with NPDR =3, Mild

NPDR = 1, Moderate NPDR = 2 according to the ETDRS Grading scheme, with 3 eyes

exhibiting macular edema and 7 which had previous treatment for DR (laser and/or IVI).

4.2.1. Deep Neural Network Algorithm Performance

An example of the automated segmentation output for both the prototype and

commercial systems is shown in Figure 4.1 along with the corresponding original image

and manual segmentation. For the images acquired with the commercial system, the

accuracy (healthy: 0.796, diabetic: 0.831), sensitivity (healthy: 0.763, diabetic: 0.758)

and specificity (healthy: 0.869, diabetic: 0.913) of the deep learning algorithm were

calculated. These measures were also calculated for the images acquired with the

1060nm prototype system: accuracy (healthy: 0.797, diabetic: 0.833), sensitivity

(healthy: 0.806, diabetic: 0.733) and specificity (healthy: 0.790, diabetic: 0.881).

37

Figure 4.1 An example image from both the prototype and commercial systems with their corresponding manual and automated segmentations. As some data within the commercial dataset contained an icon in the lower left corner a mask was applied and can be seen in the lower left corner of the automated segmentation result.

4.2.2. Clinical Outcome Measures

Representative healthy and diabetic images from both systems using both

segmentation methods for the FAZ perimeter, minimum diameter, and maximum

diameter are shown in Figure 4.2. Table 1 shows the results for the clinical outcome

measures in both systems. No significant difference existed between the means of the

clinical parameters derived from the manual and automated segmentations of images

from the OCT-A systems. All statistical measures are reported in Table 1.

38

Table 1. Clinical Outcome Measures

OCT-A System

FAZ Area (mm2)

Minimum FAZ Diameter (mm)

Maximum FAZ Diameter (mm)

FAZ Eccentricity

Perifoveal Capillary Density

Prototype OCT-A

Healthy (n = 21) Manual Automated T-test ICC

Diabetic (n = 7) Manual Automated T-test ICC

0.300 ± 0.146 0.279 ± 0.113 p = 0.11 0.956 0.415±0.164 0.366±0.104 p=0.23 0.886

0.485 ± 0.174 0.470 ± 0.130 p = 0.58 0.890 0.516 ± 0.161 0.446 ± 0.088 p=0.13 0.830

0.707 ± 0.229 0.707 ± 0.159 p = 0.97 0.869 0.928 ± 0.164 0.895 ± 0.172 p=0.52 0.859

0.721 ± 0.083 0.740 ± 0.092 p = 0.34 0.854 0.819 ± 0.083 0.852 ± 0.052 p=0.30 0.621

0.381 ± 0.043 0.388 ± 0.015 p = 0.54 0.680 0.311±0.047 0.281±0.015 p=0.09 0.634

Commercial OCT-A

Healthy (n = 31) Manual Automated T-test ICC

Diabetic (n = 12) Manual Automated T-test ICC

0.310 ± 0.122 0.314 ± 0.113 p = 0.39 0.988 0.473 ± 0.295 0.460 ± 0.262 p = 0.49 0.988

0.522 ± 0.119 0.525 ± 0.111 p = 0.66 0.984 0.549 ± 0.190 0.490 ± 0.102 p = 0.16 0.777

0.699 ± 0.132 0.715 ± 0.118 p = 0.09 0.961 0.911 ± 0.280 0.926 ± 0.294 p = 0.48 0.985

0.662 ± 0.067 0.675 ± 0.076 p = 0.20 0.828 0.784 ± 0.069 0.818 ± 0.082 p = 0.20 0.572

0.560 ± 0.049 0.560 ± 0.015 p = 0.99 0.484 0.453 ± 0.093 0.472 ± 0.047 p = 0.27 0.839

Table 1: The mean (± std) of the clinical outcome parameters (FAZ area, minimum diameter, maximum diameter, eccentricity and perifoveal capillary density) are shown for both healthy and diabetic eyes and both OCT-A systems. The p-value from the paired t-test and the Intraclass Correlation Coefficient (ICC) value is also shown to compare manual and automated methods.

For both OCT-A systems, eyes with DR had significantly lower perifoveal

capillary density (p<0.01), greater maximum diameter (p=0.04), and greater eccentricity

(p<0.01) compared to the healthy normals. There was no significant difference in FAZ

area or minimum diameter in either system.

39

Figure 4.2 FAZ perimeter (yellow), maximum diameter (green) and minimum diameter (red) shown for example healthy and diabetic data from both systems using both manually and automated segmentations. As some data within the commercial dataset contained an icon in the lower left corner a mask was applied and can be seen in the lower left corner of the commercial automated segmentation results (white) and the lower left corner in the manual segmentation of the commercial diabetic image (black).

4.3. Discussion

This study demonstrates the ability of a machine-learning based automated

segmentation algorithm to segment the vessels of both healthy and diabetic eyes

imaged with prototype and commercial OCT-A devices. The major findings of the study

are: 1) Pixel-wise, the accuracy of the automated segmentation was comparable to that

of a manual rater in both OCT-A platforms, 2) Deep learning based segmentation can

reliably quantify the perifoveal capillary density compared to manual segmentation

across the prototype and commercial OCT-A platforms and 3) Deep learning based

segmentation has the capacity to reliably quantify the FAZ area, eccentricity, maximum

and minimum diameter compared to manual raters in both OCT-A platforms.

The fovea centralis is the anatomical area responsible for the highest visual

acuity. With the exception of the foveola, the metabolic demands of the fovea centralis

are met by a unique arrangement of inner retinal capillaries. These end-artertial

capillaries lack anastomoses which make this retinal eccentricity especially vulnerable to

ischemic insult by retinal vascular diseases, including diabetic retinopathy. Perifoveal

40

capillary ischemia and FAZ enlargement are well-documented observations of macular

ischemia in diabetic retinopathy and are correlated to disease severity and progression.

While FA remains the current clinical standard for evaluating macular ischemia, its

invasive nature and potential adverse events makes it challenging to incorporate in

regular screening and frequent follow-up of patients with diabetic retinopathy. OCT-A is

an alternative non-invasive, label-free imaging modality that has been favorably

compared to histological representation and FA in in the visualization of perifoveal

circulation in healthy subjects and patients with DR. Accurate methods of quantification

and analysis of OCT-A images are of great research and clinical interest. Optovue’s

built-in automated vessel segmentation extrapolates skeletonized outputs to show

vessel density maps [67], [68]. Potential limitations to this method include

underestimating vessel density in areas with thicker vessels and decreased sensitivity to

capillary dropout [69]. A more recently published approach by Schottenhamml et al. [69]

takes advantage of a ‘vesselness’ filter to exploit the interconnective nature of the retinal

vasculature and generate more detailed vessel segmentations; however, inaccuracies

seem to result with vessel shape at vertices. The deep learning based method used in

this report takes advantage of the 2-dimensional spatial structure of training images and

was able to accurately mimic manual segmentations. Pixel-wise, the accuracy of the

automated segmentation outputs compared to manual ranged from 79-84%, which is

comparable to a reported inter-rater manual segmentation of ~83% agreement [31].

Perifoveal capillary non-perfusion is defined as the pathological enlargement of

inter-vessel distances between perifoveal capillary networks. Perifoveal capillary non-

perfusion represents an important biomarker in the pathogenesis of DR. It has been

suggested that the early perifoveal capillary non-perfusion could be present in the

absence of obvious diabetic retinopathy on clinical examination [70] OCT-A has been

shown to be able to delineate the perifoveal capillary networks precisely and consistently

in patients with DR [71]. This serves as a motivation to develop an automated tool that

can accurately and reliably quantify the perifoveal capillary density. No statistically

significant difference in the means of the perifoveal capillary densities when comparing

measurements derived from the automated and manual segmentations was found. In

comparing diabetic and healthy eyes, the automated outputs from both systems found a

significantly lower (p<0.01) perifoveal capillary density in the DR eyes. This suggests

41

that perifoveal capillary density calculated using deep learning based segmentation

might be a clinically useful tool for evaluating diabetic retinopathy.

The FAZ approximately delineates the location of the foveola within the fovea

centralis. The absence of retinal vasculature is believed to help optimize the image on

central cones. Increased FAZ area has been well-correlated with decreased visual acuity

[9] and the severity of capillary nonperfusion [6], [8] in patients with DR. Although there

is high inter-individual variability in FAZ metrics for healthy eyes, longitudinal progression

of the FAZ morphology may be a useful biomarker for DR [49]. Improved visualization of

the FAZ, enabled by OCT-A has allowed researchers to further study this area as it

relates to DR [72], retinal vein occlusion [73], and aging [67]. To assess the clinical utility

of the automated segmentations in calculating FAZ morphometric parameters, the FAZ

area, minimum and maximum diameter, and eccentricity were calculated using both the

manual and automated segmentations. No significant difference existed between the

means of the morphometric parameters derived from the manual and automated

segmentations. For both systems, the diabetic eyes were found to have a greater FAZ

maximum diameter (p=0.04), and greater eccentricity (p<0.01) compared to the healthy

normals while no no significant difference in FAZ area or minimum diameter was noted.

This lack of correlation is likely due to the high inter-individual variability of FAZ metrics

and low number of severe NPDR subjects in the study. A non-circular FAZ shape may

be a more reliable biomarker as indicated by the greater eccentricity in DR eyes.

This study demonstrates the ability of a deep learning based automated

segmentation algorithm to reliably segment the perifoveal microvasculature and provide

clinically useful FAZ morphological measures. A limitation of this study is a restricted

sample size for the diabetic groups. Additionally, as the performance of a deep learning

based approach is limited by the quality of the training data, the automated

segmentation performance is limited by image quality and manual vessel segmentations.

Common OCT-A image quality issues such as low signal-to-noise ratio within the FAZ

and motion artefacts, as seen in Figure 4.3, caused erroneous segmentations in some

cases. Although the manual segmentations were reviewed by two trained raters in an

attempt to mitigate manual segmentation error, the segmentations were reviewed on a

holistic level whereas the machine learns on a pixel-by-pixel basis. Another limitation is

the need for a database of segmented images for each field of view and each machine.

The commercial machine allows the imaging of three different fields of view (3x3, 6x6,

42

8x8mm), of which only one (3x3mm) was chosen to analyze in this study.

Experimentally, performance variability in the vessel thickness occurred when the

training set and the segmented images were from different fields of view, therefore a

training set is needed for each field of view.

Figure 4.3 Examples of low quality input data and the automated segmentation. Due to the low signal-to-noise ratio within the FAZ, some areas were erroneously segmented. Additionally, a horizontal motion artefact can be seen cutting through the FAZ which was also incorrectly segmented in areas.

Systematic screening of people with diabetes has been shown to be a cost-

effective approach for identifying potential vision loss [74]. OCT-A is a promising new

technology that has the potential to help guide earlier management decisions and

prognosis. Deep learning automated segmentation of OCT-A may be suitable for both

commercial and research purposes for better quantification of the retinal circulation in

healthy subjects and in subjects with retinal vascular disease.

4.4. Summary

In this chapter, the use of DNNs for automatically analyzing OCT-A images was

evaluated. Using DNNs partially eliminates the need for manual segmentations.

Although a certain number of initial segmentations would be necessary to train the

machine, once the machine is properly trained manual segmentations would not be

needed for future studies. The next chapter discusses possible future studies which

could build upon the work presented in this report.

43

Chapter 5. Future Work

While techniques were presented in this report which help to overcome the

image quality and automated segmentation needs discussed in Chapter 2, more can be

done to advance the clinical utility of OCT and its derivatives. Some possible

suggestions are listed below.

5.1. 4D Registered OCTA Volumes

In order to see the temporal variations of the microvasculature, the individual

registered volumes can be evaluated separately instead of averaging. In order to

achieve efficient blood flow distribution, vascular shunting can occur which allows the

blood to bypass the capillaries in certain areas. For example, in response to cold

conditions some shunt vessels dilate to cut off blood flow to the extremities thereby

preventing heat loss. Conversely, when exercising shunt vessels to the muscles

constrict which pushes blood through the capillary networks where it can deliver oxygen

to the muscles that require it the most. This autoregulation happens within eyes as well.

By registering the networks, we can observe temporal changes within the eye and

perhaps study certain retinal vascular diseases from a different angle.

5.2. Registration of Photoreceptor Images

The strip-based registration technique described in Chapter 3 is not limited to the

registration of retinal blood vessels. In theory, we should also be able to use the pipeline

to register other retinal layers, including the photoreceptors. Figure 5.1 shows two

representative images acquired with our Adaptive Optics (AO) OCT system, which

provided a high lateral resolution enabling visualization of the photoreceptor mosaic.

44

Figure 5.1 Two AO-OCT images of photoreceptors acquired from the same patient in a similar area. Scale bar is 50μm.

By averaging these images we should be able to see more circular photoreceptors, and

in the case of under sampling, we could potentially be able to resolve features we

couldn’t see with just one volume. The Retinal Pigment Epithelium (RPE) is a layer of

the retina which researchers have not yet been able to readily visualize without the

averaging of multiple frames [75].

5.3. Apply the DNN Framework to Averaged Images

Using the framework in Chapter 3, capillaries were more readily visualized. This

improvement in CNR and SNR would improve manual rater confidence. As manual

segmentations are the input to our DNN automated segmentation tool, an improvement

in manual segmentations could result in improved and more accurate automated

segmentations.

Additionally, if given enough information the DNN could highlight other areas of

interest in the OCT-A angiograms, like microaneurysms. Microaneurysms are found in

diabetic patients and keeping track of the number of microaneurysms that appear in the

macula can be useful in tracking disease progression.

Another exciting possibility is that DNN could help us to visualize these networks

in three dimensions. As OCT-A is acquired in 3D, it is possible to obtain a volumetric

view of the vasculature. Instead of segmenting the layers of the retina and extracting 2D

images of the networks, we could obtain volumetric manual segmentations of the

45

vasculature on which the DNN could train. Manual segmenters could use either the B-

scan or C-scan view in order to properly mark the vessels. Delineating the vessels using

this technique would also eliminate the need for projection artefact removal as the

computer would be able to learn what is and is not artefact through the segmentations.

As the retinal plexus layers aren’t completely separate, this would allow clinicians and

researchers to be able to interact with the data in a new way.

46

References

[1] P. E. Z. Tan, C. Balaratnasingam, J. Xu, Z. Mammo, S. X. Han, P. Mackenzie, A. W. Kirker, D. Albiani, A. B. Merkur, M. V. Sarunic, and D.-Y. Y. Yu, “Quantitative Comparison of Retinal Capillary Images Derived By Speckle Variance Optical Coherence Tomography With Histology.,” Investig. Ophthalmol. Vis. Sci., vol. 56, no. 6, pp. 3989–3996, 2015.

[2] J. W. Y. Yau, S. L. Rogers, R. Kawasaki, E. L. Lamoureux, J. W. Kowalski, T. Bek, S.-J. Chen, J. M. Dekker, A. Fletcher, J. Grauslund, S. Haffner, R. F. Hamman, M. K. Ikram, T. Kayama, B. E. K. Klein, R. Klein, S. Krishnaiah, K. Mayurasakorn, J. P. O’Hare, T. J. Orchard, M. Porta, M. Rema, M. S. Roy, T. Sharma, J. Shaw, H. Taylor, J. M. Tielsch, R. Varma, J. J. Wang, N. Wang, S. West, L. Xu, M. Yasuda, X. Zhang, P. Mitchell, and T. Y. Wong, “Global prevalence and major risk factors of diabetic retinopathy,” Diabetes Care, vol. 35, no. 3, pp. 556–564, 2012.

[3] S. Resnikoff, D. Pascolini, D. Etya’ale, I. Kocur, R. Pararajasegaram, G. P. Pokharel, and S. P. Mariotti, “Global data on visual impairment in the year 2002.,” Bull. World Health Organ., vol. 82, no. 11, pp. 844–51, 2004.

[4] M. M. Nentwich and M. W. Ulbig, “Diabetic retinopathy - ocular complications of diabetes mellitus.,” World J. Diabetes, vol. 6, no. 3, pp. 489–99, 2015.

[5] D. A. Antonetti, R. Klein, and T. W. Gardner, “Diabetic Retinopathy,” N Engl J Med, vol. 366, pp. 1227–1239, 2012.

[6] G. Bresnick, G. De Venecia, F. Myers, J. Harris, and M. Davis, “Retinal ischemia in diabetic retinopathy,” Arch. Ophthalmol., vol. 93, no. 12, pp. 1300–1310, 1975.

[7] G. A. Lutty, “Effects of diabetes on the eye,” Investig. Ophthalmol. Vis. Sci., vol. 54, no. 14, 2013.

[8] O. Arend, S. Wolf, A. Harris, and M. Reim, “The Relationship of Macular Microcirculation to Visual-Acuity in Diabetic-Patients,” Arch. Ophthalmol., vol. 113, no. 5, pp. 610–614, 1995.

[9] C. Balaratnasingam, M. Inoue, S. Ahn, J. McCann, E. Dhrami-Gavazi, L. A. Yannuzzi, and K. B. Freund, “Visual Acuity Is Correlated with the Area of the Foveal Avascular Zone in Diabetic Retinopathy and Retinal Vein Occlusion,” Ophthalmology, 2016.

[10] Y. C. Tham, X. Li, T. Y. Wong, H. A. Quigley, T. Aung, and C. Y. Cheng, “Global prevalence of glaucoma and projections of glaucoma burden through 2040: A systematic review and meta-analysis,” Ophthalmology, vol. 121, no. 11, pp. 2081–2090, 2014.

47

[11] H. A. Quigley and A. T. Broman, “The number of people with glaucoma worldwide in 2010 and 2020.,” Br. J. Ophthalmol., vol. 90, no. 3, pp. 262–267, Mar. 2006.

[12] Y. H. Kwon, J. H. Fingert, M. H. Kuehn, and W. L. M. Alward, “Primary open-angle glaucoma.,” N. Engl. J. Med., vol. 360, no. 11, pp. 1113–1124, 2009.

[13] E. J. Lee, T.-W. Kim, M. Kim, and H. Kim, “Influence of lamina cribrosa thickness and depth on the rate of progressive retinal nerve fiber layer thinning.,” Ophthalmology, vol. 122, no. 4, pp. 721–729, Apr. 2015.

[14] A. Giangiacomo, D. Garway-Heath, and J. Caprioli, “Diagnosing glaucoma progression: current practice and promising technologies.,” Curr. Opin. Ophthalmol., vol. 17, no. 2, pp. 153–162, Apr. 2006.

[15] D.-Y. Yu, S. J. Cringle, C. Balaratnasingam, W. H. Morgan, P. K. Yu, and E.-N. Su, “Retinal ganglion cells: Energetics, compartmentation, axonal transport, cytoskeletons and vulnerability.,” Prog. Retin. Eye Res., vol. 36, pp. 217–246, Sep. 2013.

[16] A. L. Kornzweig, I. Eliasoph, and M. Feldstein, “Selective atrophy of the radial peripapillary capillaries in chronic glaucoma.,” Arch. Ophthalmol. (Chicago, Ill. 1960), vol. 80, no. 6, pp. 696–702, Dec. 1968.

[17] P. Henkind, “Radial peripapillary capillaries of the retina. I. Anatomy: human and comparative.,” Br. J. Ophthalmol., vol. 51, no. 2, pp. 115–123, Feb. 1967.

[18] P. Henkind, “Symposium on glaucoma: joint meeting with the National Society for the Prevention of Blindness. New observations on the radial peripapillary capillaries.,” Invest. Ophthalmol., vol. 6, no. 2, pp. 103–108, Apr. 1967.

[19] P. K. Yu, S. J. Cringle, and D.-Y. Yu, “Correlation between the radial peripapillary capillaries and the retinal nerve fibre layer in the normal human retina.,” Exp. Eye Res., vol. 129, pp. 83–92, Dec. 2014.

[20] R. F. Spaide, J. M. J. Klancnik, and M. J. Cooney, “Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography.,” JAMA Ophthalmol., vol. 133, no. 1, pp. 45–50, Jan. 2015.

[21] M. S. Mahmud, D. W. Cadotte, B. Vuong, C. Sun, T. W. H. Luk, A. Mariampillai, and V. X. D. Yang, “Review of speckle and phase variance optical coherence tomography to visualize microvascular networks.,” J. Biomed. Opt., vol. 18, no. 5, p. 50901, 2013.

[22] A. Zhang, Q. Zhang, C.-L. Chen, and R. K. Wang, “Methods and algorithms for optical coherence tomography-based angiography: a review and comparison,” J. Biomed. Opt., vol. 20, no. 10, p. 100901, 2015.

48

[23] I. Gorczynska, J. V. Migacz, R. J. Zawadzki, A. G. Capps, and J. S. Werner, “Comparison of amplitude-decorrelation, speckle-variance and phase-variance OCT angiography methods for imaging the human retina and choroid,” Biomed. Opt. Express, vol. 7, no. 3, p. 911, 2016.

[24] Y. Jia, O. Tan, J. Tokayer, B. Potsaid, Y. Wang, J. J. Liu, M. F. Kraus, H. Subhash, J. G. Fujimoto, J. Hornegger, and D. Huang, “Split-spectrum amplitude-decorrelation angiography with optical coherence tomography.,” Opt. Express, vol. 20, no. 4, pp. 4710–25, 2012.

[25] J. Xu, S. Han, C. Balaratnasingam, Z. Mammo, K. S. K. Wong, S. Lee, M. Cua, M. Young, A. Kirker, D. Albiani, F. Forooghian, P. Mackenzie, A. Merkur, D.-Y. Yu, and M. V. Sarunic, “Retinal angiography with real-time speckle variance optical coherence tomography,” Br. J. Ophthalmol., pp. 1–5, 2015.

[26] Z. Mammo, C. Balaratnasingam, P. Yu, J. Xu, M. Heisler, P. Mackenzie, A. Merkur, A. Kirker, D. Albiani, K. B. Freund, M. V. Sarunic, D.-Y. Y. Yu, K. Bailey Freund, M. V. Sarunic, and D.-Y. Y. Yu, “Quantitative Noninvasive Angiography of the Fovea Centralis Using Speckle Variance Optical Coherence TomographySpeckle Variance Optical Coherence Tomography of Macula,” Investig. Ophthalmol. Vis. Sci., vol. 56, no. 9, pp. 5074–5086, 2015.

[27] K. R. Mendis, C. Balaratnasingam, P. Yu, C. J. Barry, I. L. McAllister, S. J. Cringle, and D.-Y. Yu, “Correlation of histologic and clinical images to determine the diagnostic value of fluorescein angiography for studying retinal capillary detail.,” Invest. Ophthalmol. Vis. Sci., vol. 51, no. 11, pp. 5864–5869, Nov. 2010.

[28] G. Chan, C. Balaratnasingam, J. Xu, Z. Mammo, S. Han, P. Mackenzie, A. Merkur, A. Kirker, D. Albiani, M. V. Sarunic, and D.-Y. Yu, “In vivo optical imaging of human retinal capillary networks using speckle variance optical coherence tomography with quantitative clinico-histological correlation,” Microvasc. Res., vol. 100, pp. 32–39, 2015.

[29] P. K. Yu, C. Balaratnasingam, J. Xu, W. H. Morgan, Z. Mammo, S. Han, P. Mackenzie, A. Merkur, A. Kirker, D. Albiani, M. V. Sarunic, and D.-Y. Yu, “Label-Free Density Measurements of Radial Peripapillary Capillaries in the Human Retina,” PLoS One, vol. 10, no. 8, p. e0135151, 2015.

[30] Z. Mammo, M. Heisler, C. Balaratnasingam, S. Lee, D. Y. Yu, P. Mackenzie, S. Schendel, A. Merkur, A. Kirker, D. Albiani, E. Navajas, M. F. Beg, W. Morgan, and M. V. Sarunic, “Quantitative Optical Coherence Tomography Angiography of Radial Peripapillary Capillaries in Glaucoma, Glaucoma Suspect, and Normal Eyes,” Am. J. Ophthalmol., vol. 170, pp. 41–49, 2016.

[31] P. Prentasic, M. Heisler, S. Lee, Z. Mammo, A. B. Merkur, E. Navajas, M. F. Beg, M. V. Sarunic, and S. Loncaric, “Segmentation of the Foveal Microvasculature Using Deep Learning Networks,” J. Biomed. Opt., vol. 21, no. 7, 2016.

49

[32] M. Heisler, F. Chan, Z. Mammo, C. Balaratnasingam, P. Prentasic, G. Docherty, S. Rajapakse, S. Lee, A. Merkur, A. Kirker, D. Albiani, D. Maberley, K. B. Freund, M. F. Beg, S. Loncaric, M. V. Sarunic, and E. V Navajas, “Deep Learning Vessel Segmentation and Quantification of the Foveal Vascular Zone using Prototype and Commercial OCT-A Platforms,” Investig. Ophthalmol. Vis. Sci. Submitted 2017

[33] M. Heisler, S. Lee, Z. Mammo, Y. Jian, M. Ju, A. Merkur, E. Navajas, C. Balaratnasingam, M. F. Beg, and M. V. Sarunic, “Strip-based registration of serially acquired optical coherence tomography angiography Strip-based registration of serially acquired optical,” J. Biomed. Opt., vol. 22, no. 3, p. 36007, 2017.

[34] C. E. Pang, E. V Navajas, S. J. Warner, M. Heisler, and M. V Sarunic, “Acute Macular Neuroretinopathy Associated With Chikungunya Fever,” Ophthalmic Surg Lasers Imaging Retin., vol. 47, no. 6, pp. 596–9, 2016.

[35] Y. Jian, S. Lee, M. J. Ju, M. Heisler, W. Ding, R. J. Zawadzki, S. Bonora, and M. V Sarunic, “Lens-based wavefront sensorless adaptive optics swept source OCT,” Sci. Rep., vol. 6:27620, pp. 1–10, 2016.

[36] H. R. G. W. Verstraete, M. Heisler, M. J. Ju, D. Wahl, L. Bliek, J. Kalkman, S. Bonora, Y. Jian, M. Verhaegen, and M. V Sarunic, “Wavefront sensorless adaptive optics OCT with the DONE algorithm for in vivo human retinal imaging,” Biomed. Opt. Express, vol. 8, no. 4, pp. 2261–2275, Apr. 2017.

[37] M. Heisler, W. L. Quong, S. Lee, S. Han, M. F. Beg, M. V Sarunic, and P. J. Mackenzie, “Anterior Segment Optical Coherence Tomography for Targeted Transconjunctival Suture Placement in Overfiltering Trabeculectomy Blebs,” J. Glaucoma, vol. 0, no. 0, pp. 1–5, 2017.

[38] S. Lee, M. Heisler, P. J. Mackenzie, M. V Sarunic, and M. Faisal, “Quantifying Variability in Longitudinal Peripapillary RNFL and Choroidal Layer Thickness Using Surface Based Registration of OCT Images,” Transl. Vis. Sci. Technol., vol. 6, no. 1, pp. 1–20, 2017.

[39] S. Lee, M. Heisler, K. Popuri, N. Charon, A. Trouvé, P. J. Mackenzie, M. V. Sarunic, and M. F. Beg, “Age and glaucoma-related changes in retinal nerve fiber layer and choroid: point-wise analysis and visualization using functional shapes registration,” Front. Neurosci. Submitted 2017

[40] A. Mariampillai, B. A. Standish, E. H. Moriyama, M. Khurana, N. R. Munce, M. K. K. Leung, J. Jiang, A. Cable, B. C. Wilson, I. A. Vitkin, and V. X. D. Yang, “Speckle variance detection of microvasculature using swept-source optical coherence tomography.,” Opt. Lett., vol. 33, no. 13, pp. 1530–1532, 2008.

50

[41] Y. Jian, K. Wong, and M. V Sarunic, “Graphics processing unit accelerated optical coherence tomography processing at megahertz axial scan rate and high resolution video rate volumetric rendering.,” J. Biomed. Opt., vol. 18, no. 2, p. 26002, 2013.

[42] J. Xu, K. Wong, Y. Jian, and M. V Sarunic, “Real-time acquisition and display of flow contrast using speckle variance optical coherence tomography in a graphics processing unit.,” J. Biomed. Opt., vol. 19, no. 2, p. 26001, 2014.

[43] S. Lee, N. Fallah, F. Forooghian, A. Ko, K. Pakzad-Vaezi, A. B. Merkur, A. W. Kirker, D. a. Albiani, M. Young, M. V. Sarunic, and M. F. Beg, “Comparative analysis of repeatability of manual and automated choroidal thickness measurements in nonneovascular age-related macular degeneration,” Investig. Ophthalmol. Vis. Sci., vol. 54, no. 4, pp. 2864–71, 2013.

[44] J. G. Christenbury, M. A. Klufas, T. C. Sauer, and D. Sarraf, “OCT Angiography of Paracentral Acute Middle Maculopathy Associated With Central Retinal Artery Occlusion and Deep Capillary Ischemia,” Ophthalmic Surgery, Lasers Imaging Retin., vol. 46, no. 5, pp. 579–581, 2015.

[45] M. A. Khan, E. Rahimy, A. Shahlaee, J. Hsu, and A. C. Ho, “En face optical coherence tomography imaging of deep capillary plexus abnormalities in paracentral acute middle maculopathy,” Ophthalmic Surg. Lasers Imaging Retin., vol. 46, no. 9, pp. 972–975, 2015.

[46] G. Casalino, M. Williams, C. McAvoy, F. Bandello, and U. Chakravarthy, “Optical coherence tomography angiography in paracentral acute middle maculopathy secondary to central retinal vein occlusion,” Eye, vol. 44, no. January, pp. 1–6, 2016.

[47] N. Hasegawa, M. Nozaki, N. Takase, M. Yoshida, and Y. Ogura, “New Insights Into Microaneurysms in the Deep Capillary Plexus Detected by Optical Coherence Tomography Angiography in Diabetic Macular Edema.,” Invest. Ophthalmol. Vis. Sci., vol. 57, no. 9, p. OCT348-55, Jul. 2016.

[48] P. D. Bradley, D. A. Sim, P. A. Keane, J. Cardoso, R. Agrawal, A. Tufail, and C. A. Egan, “The Evaluation of Diabetic Macular Ischemia Using Optical Coherence Tomography Angiography.,” Invest. Ophthalmol. Vis. Sci., vol. 57, no. 2, pp. 626–631, Feb. 2016.

[49] T. S. Hwang, S. S. Gao, L. Liu, A. K. Lauer, S. T. Bailey, C. J. Flaxel, D. J. Wilson, D. Huang, and Y. Jia, “Automated Quantification of Capillary Nonperfusion Using Optical Coherence Tomography Angiography in Diabetic Retinopathy.,” JAMA Ophthalmol., vol. 134, no. 4, pp. 367–73, 2016.

[50] K. Kurokawa, K. Sasaki, S. Makita, Y.-J. Hong, and Y. Yasuno, “Three-dimensional retinal and choroidal capillary imaging by power Doppler optical coherence angiography with adaptive optics.,” Opt. Express, vol. 20, no. 20, pp. 22796–22812, Sep. 2012.

51

[51] M. F. Kraus, J. J. Liu, J. Schottenhamml, C.-L. Chen, A. Budai, L. Branchini, T. Ko, H. Ishikawa, G. Wollstein, J. S. Schuman, J. S. Duker, J. G. Fujimoto, and J. Hornegger, “Quantitative 3D-OCT motion correction with tilt and illumination correction, robust similarity measure and regularization.,” Biomed. Opt. Express, vol. 5, no. 8, pp. 2591–613, 2014.

[52] H. C. Hendargo, R. Estrada, S. J. Chiu, C. Tomasi, S. Farsiu, and J. A. Izatt, “Automated non-rigid registration and mosaicing for robust imaging of distinct retinal capillary beds using speckle variance optical coherence tomography.,” Biomed. Opt. Express, vol. 4, no. 6, pp. 803–21, 2013.

[53] P. Zang, G. Liu, M. Zhang, C. Dongye, J. Wang, A. D. Pechauer, T. S. Hwang, D. J. Wilson, D. Huang, D. Li, and Y. Jia, “Automated motion correction using parallel- strip registration for wide-field en face OCT angiogram,” Biomed. Opt. Express, vol. 7, no. 7, pp. 3822–3832, 2016.

[54] D. Ruminski, B. L. Sikorski, D. Bukowska, M. Szkulmowski, K. Krawiec, G. Malukiewicz, L. Bieganowski, and M. Wojtkowski, “OCT angiography by absolute intensity difference applied to normal and diseased human retinas.,” Biomed. Opt. Express, vol. 6, no. 8, pp. 2738–2754, Aug. 2015.

[55] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004.

[56] A. Zhang, Q. Zhang, and R. K. Wang, “Minimizing projection artifacts for accurate presentation of choroidal neovascularization in OCT micro-angiography,” Biomed. Opt. Express, vol. 6, no. 10, pp. 369–380, 2015.

[57] J. P. (Industrial L. & M. Lewis, “Fast Normalized Cross-Correlation Template Matching by Cross-,” Vis. Interface, vol. 1995, no. 1, pp. 1–7, 1995.

[58] A. Ozcan, A. Bilenca, A. E. Desjardins, B. E. Bouma, and G. J. Tearney, “Speckle reduction in optical coherence tomography images using digital filtering.,” J. Opt. Soc. Am. A. Opt. Image Sci. Vis., vol. 24, no. 7, pp. 1901–1910, 2007.

[59] D. C. Adler, T. H. Ko, and J. G. Fujimoto, “Speckle reduction in optical coherence tomography images by use of a spatially adaptive wavelet filter.,” Opt. Lett., vol. 29, no. 24, pp. 2878–2880, 2004.

[60] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity.,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.

[61] S. G. K. Gadde, N. Anegondi, D. Bhanushali, L. Chidambara, N. K. Yadav, A. Khurana, and A. S. Roy, “Quantification of vessel density in retinal optical coherence tomography angiography images using local fractal dimension,” Investig. Ophthalmol. Vis. Sci., vol. 57, no. 1, pp. 246–252, 2016.

52

[62] T. E. de Carlo, A. Romano, N. K. Waheed, and J. S. Duker, “A review of optical coherence tomography angiography (OCTA),” Int. J. Retin. Vitr., vol. 1, no. 1, p. 5, 2015.

[63] Y. Jia, J. C. Morrison, J. Tokayer, O. Tan, L. Lombardi, B. Baumann, C. D. Lu, W. Choi, J. G. Fujimoto, and D. Huang, “Quantitative OCT angiography of optic nerve head blood flow,” Biomed. Opt. Express, vol. 3, no. 12, pp. 3127–3137, Dec. 2012.

[64] D. Huang, Y. Jia, S. S. Gao, B. Lumbroso, and M. Rispoli, “Optical Coherence Tomography Angiography Using the Optovue Device.,” Dev. Ophthalmol., vol. 56, pp. 6–12, 2016.

[65] Early Treatment Diabetic Retinopathy Study Research Group, “Fluorescein angiographic risk factors for progression of diabetic retinopathy. ETDRS report number 13. Early Treatment Diabetic Retinopathy Study Research Group.,” Ophthalmology, vol. 98, no. 5 Suppl, pp. 834–40, 1991.

[66] Optovue Incorporated, AngioVue Imaging System Brochure. Fremont, CA, 2015.

[67] N. A. Iafe, N. Phasukkijwatana, X. Chen, and D. Sarraf, “Retinal Capillary Density and Foveal Avascular Zone Area Are Age-Dependent: Quantitative Analysis Using Optical Coherence Tomography AngiographyQuantitative Analysis Using OCT Angiography,” Invest. Ophthalmol. Vis. Sci., vol. 57, no. 13, p. 5780, 2016.

[68] S. a. Agemy, N. K. Scripsema, C. M. Shah, T. Chui, P. M. Garcia, J. G. Lee, R. C. Gentile, Y.-S. Hsiao, Q. Zhou, T. Ko, and R. B. Rosen, “RETINAL VASCULAR PERFUSION DENSITY MAPPING USING OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY IN NORMALS AND DIABETIC RETINOPATHY PATIENTS.,” Retina, vol. 35, no. 11, pp. 2353–2363, Nov. 2015.

[69] J. Schottenhamml, E. M. Moult, S. Ploner, B. Lee, E. A. Novais, E. Cole, S. Dang, C. D. Lu, L. Husvogt, N. K. Waheed, J. S. Duker, J. Hornegger, and J. G. Fujimoto, “AN AUTOMATIC, INTERCAPILLARY AREA-BASED ALGORITHM FOR QUANTIFYING DIABETES-RELATED CAPILLARY DROPOUT USING OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY,” Retina, p. , 2016.

[70] T. Nagaoka, E. Sato, A. Takahashi, H. Yokota, K. Sogawa, and A. Yoshida, “Impaired retinal circulation in patients with type 2 diabetes mellitus: retinal laser Doppler velocimetry study.,” Invest. Ophthalmol. Vis. Sci., vol. 51, no. 12, pp. 6729–6734, Dec. 2010.

[71] S. DA, de Carlo TE, A. M, and et al, “Select features of diabetic retinopathy on swept-source optical coherence tomographic angiography compared with fluorescein angiography and normal eyes,” JAMA Ophthalmol., vol. 134, no. 6, pp. 644–650, 2016.

53

[72] M. Al-Sheikh, H. Akil, M. Pfau, and S. R. Sadda, “Swept-Source OCT Angiography Imaging of the Foveal Avascular Zone and Macular Capillary Network Density in Diabetic RetinopathyOCT-Angiography in Diabetic Retinopathy,” Invest. Ophthalmol. Vis. Sci., vol. 57, no. 8, p. 3907, 2016.

[73] J.-W. Kang, R. Yoo, Y. H. Jo, and H. C. Kim, “CORRELATION OF MICROVASCULAR STRUCTURES ON OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY WITH VISUAL ACUITY IN RETINAL VEIN OCCLUSION.,” RETINA, vol. Publish Ah, Nov. 9000.

[74] S. Philip, A. D. Fleming, K. A. Goatman, S. Fonseca, P. Mcnamee, G. S. Scotland, G. J. Prescott, P. F. Sharp, and J. A. Olson, “The efficacy of automated ‘disease/no disease’ grading for diabetic retinopathy in a systematic screening programme,” Br. J. Ophthalmol., vol. 91, no. 11, pp. 1512–1517, Nov. 2007.

[75] Z. Liu, O. P. Kocaoglu, and D. T. Miller, “3D Imaging of Retinal Pigment Epithelial Cells in the Living Human Retina3D Imaging of RPE Cells in Living Human Retina,” Invest. Ophthalmol. Vis. Sci., vol. 57, no. 9, p. OCT533, 2016.

54

Appendix. Further Examples of Averaged OCT-A Images

Figure A.1 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

55

Figure A.2 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

Figure A.3 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

56

Figure A.4 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

Figure A.5 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

57

Figure A.6 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

Figure A.7 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.

58

Figure A.8 Template image, mean, and median images (all retinal layers, superficial and deep plexus) for Subject 6 OS, a healthy male subject, 56 years of age.