Quantitative optical coherence tomography angiography: A review

Abstract

As a new optical coherence tomography (OCT) modality, OCT angiography (OCTA) provides a noninvasive method to detect microvascular distortions correlated with eye conditions. By providing unparalleled capability to differentiate individual plexus layers in the retina, OCTA has demonstrated its excellence in clinical management of diabetic retinopathy, glaucoma, sickle cell retinopathy, diabetic macular edema, and other eye diseases. Quantitative OCTA analysis of retinal and choroidal vasculatures is essential to standardize objective interpretations of clinical outcome. Quantitative features, including blood vessel tortuosity, blood vessel caliber, blood vessel density, vessel perimeter index, fovea avascular zone area, fovea avascular zone contour irregularity, vessel branching coefficient, vessel branching angle, branching width ratio, and choroidal vascular analysis have been established for objective OCTA assessment. Moreover, differential artery–vein analysis has been recently demonstrated to improve OCTA performance for objective detection and classification of eye diseases. In this review, technical rationales and clinical applications of these quantitative OCTA features are summarized, and future prospects for using these quantitative OCTA features for artificial intelligence classification of eye conditions are discussed.

Impact statement

OCT angiography (OCTA) provides a noninvasive method to detect microvascular distortions correlated with eye conditions. Quantitative analysis of OCTA is essential to standardize objective interpretations of clinical outcome. This review summarizes technical rationales and clinical applications of quantitative OCTA features.

Introduction

As one part of the central nervous system, the retina is a neurovascular complex network located at the back of the eye. Color fundus photography has provided valuable information for eye disease detection and treatment assessment, but the spatial resolution and image contrast are limited to reveal subtle distortions in early stages of eye diseases. Scanning laser ophthalmoscopy^1,2 and adaptive optics^3–5 imaging systems provide enhanced image resolution, and fundus angiography^6,7 allows better contrast of retinal vasculatures. However, these imaging approaches lack sectioning capability to differentiate individual retinal neural layers and vascular plexuses. It is known that different diseases and stages can target retinal neurons and vasculatures in different ways. Given the unprecedented capability to differentiate individual functional layers, optical coherence tomography (OCT)⁸ has been extensively employed for depth-resolved examination of morphological abnormalities caused by eye diseases.^9–11

As a new OCT modality, OCT angiography (OCTA) provides a noninvasive method to differentiate individual plexus layers in the retina.^12,13 Since its first commercial product in 2014, OCTA has quickly demonstrated its excellence in clinical management of diabetic retinopathy (DR),^14,15 glaucoma,^16,17 sickle cell retinopathy (SCR),¹⁸ age-related macular degeneration (AMD),¹⁹ and other eye diseases. Quantitative OCTA analysis is essential to standardize objective interpretation of clinical outcomes. Multiple OCTA features have been recently developed for quantitative analysis of vascular distortions due to eye conditions. In the following sections, technical rationales of these quantitative OCTA features will be summarized. Current status and future prospects of using OCTA features for objective detection and artificial intelligence (AI) classification of eye diseases will be discussed.

Quantitative OCTA features

In this section, technical rationale of OCTA features, i.e. blood vessel tortuosity (BVT), blood vessel caliber (BVC), blood vessel density (BVD), vessel perimeter index (VPI), fovea avascular zone area (FAZ-A), FAZ contour irregularity (FAZ-CI), vessel complexity index (VCI), branchpoint analysis (BPA), differential artery–vein (A–V) analysis, flow analysis, and choroidal neurovasculature (CNV) analysis will be explained sequentially. To help the explanation of quantitative OCTA analysis, Figure 1 illustrates major procedures of quantitative feature extraction.

Figure 1.

Feature extraction for quantitative OCTA analysis. (a) Representative OCTA image from a DR patient. (b) Segmented blood vessel map. (c) Skeletonized blood vessel map (red) with segmented fovea (blue region) and FAZ contour (green curve). One representative vessel branch is highlighted in green with X and Y endpoints identified with yellow dots. (d) Vessel perimeter map. (e) FD contour map. Source: Modified from Alam et al.²⁰ (A color version of this figure is available in the online journal.)

BVD

BVD, also named as vessel density (VD),²¹ vessel area density,²² capillary density,²³ or percent area of non-perfusion,²⁴ reflects the ratio of the image area occupied by the blood vessels (Figure 1(b)). Eye diseases such as DR,^25,26 SCR,^27,28 AMD,^29,30 glaucoma,^31,32 and vein occlusion (VO)^33,34 may involve vessel abnormalities, including ischemia and drop out zones in retinal and choroidal vasculatures. Most of these diseases manifest at the capillary level at early phases, which can be detected in OCTA. BVD can be quantified as³⁵

$$\mathit{BVD}=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n}A(x,y)}{{{\displaystyle \sum }}_{x=1,y=1}^{n}I(x,y)}$$ (1)

where A(x, y) represents the pixels occupied by the vessels, and I(x, y) represents all the pixels in the OCTA image. In this article, the OCTA image is assumed to be a square frame, consisting of n × n pixels, and x and y correspond to the horizontal and vertical coordinates of individual pixels.

If the skeletonized vessel map (Figure 1(c)) is used for VD analysis, this feature is alternatively termed as vessel skeleton density (VSD) or skeleton density.³² The VSD can be quantified as³⁶

$$\mathit{VSD}=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n}S(x,y)}{{{\displaystyle \sum }}_{x=1,y=1}^{n}I(x,y)}$$ (2)

where S(x, y) represents the pixels occupied by the vessel skeleton and I(x, y) represents all the pixels in the OCTA.

BVC

BVC, also named as vessel diameter, vessel width, or vessel diameter index,³⁷ is used to quantify vascular dilation or shrinkage due to eye conditions. BVC distortions have been commonly observed in different retinopathies, such as SCR^35,38 and DR.³⁹ The BVC can be calculated as the ratio of the vessel area to the vessel length³⁵

$$\mathit{BVC}=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n}A(x,y)}{{{\displaystyle \sum }}_{x=1,\text{y}=1}^{n}S(x,y)}$$ (3)

where A(x, y) represents the pixels occupied by the vessels in the segmented vessel map (Figure 1(b)) and S(x, y) represents the pixels occupied by the vessels in the skeletonized vessel map (Figure 1(c)). In the skeletonized map, the width of each vessel is one pixel. Therefore, ${\displaystyle {\sum }_{x=1,y=1}^{n}A(x,y)}$ represents total vessel area, and ${\displaystyle {\sum }_{x=1,y=1}^{n}S(x,y)}$ represents overall vessel length.

BVT

BVT^35,38 is a measure of the degree of vessel distortion. In normal condition, the blood vessels transport blood efficiently, with a relatively smooth structure. However, in diseased conditions,^35,38–41 the transportation efficiency of some blood vessels may be compromised due to distorted structure. The tortuosity of each blood vessel can be quantified by measuring the ratio of the geodesic distance to the Euclidean distance³⁵

$$\mathit{BVT}=\frac{1}{n}{\displaystyle \sum }_{i=1}^n\left(\frac{\left\{\begin{array}{c}G\mathit{eodesic}\mathit{distance}\mathit{between}\mathit{two}\\ \mathit{endpoints}ofa\mathit{vessel}\mathit{branch}i\end{array}\right\}}{\left\{\begin{array}{c}E\mathit{uclidean}\mathit{distance}\mathit{between}\mathit{two}\\ \mathit{endpoints}ofa\mathit{vessel}\mathit{branch}i\end{array}\right\}}\right)$$ (4)

where i represents the ith branch and n is the total number of branches; Euclidian distance is the straight-line distance between two end points of a vessel branch and geodesic distance is the total curve length between two end points (shown in Figure 1(c)).

VPI

VPI⁴² measures the ratio between overall contour length of blood vessel boundaries (Figure 1(d)) and total blood vessel area in the segmented vessel map (Figure 1(b)). VPI can reflect vessel dropout or early ischemia,^39,43 and has been used to quantify OCTA images of DR⁴³ and SCR.³⁸ It can be measured as follows³⁵

$$\mathit{VPI}=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n}P(x,y)}{{{\displaystyle \sum }}_{x=1,y=1}^{n}I(x,y)}$$ (5)

where P(x, y) represents the pixels within the vessel perimeters, i.e. the overall contour length of blood vessel boundaries (Figure 1(d)), and I(x, y) represents all the pixels occupied by the blood vessels, i.e. total blood vessel area (Figure 1(b)).

FAZ

Foveal shape can be affected by eye diseases such as DR,⁴⁴ SCR,³⁸ and VO,⁴⁵ due to parafoveal vessel drop out and foveal ischemia. FAZ-A has been demonstrated as a sensitive feature to differentiate severities of non-proliferative diabetic retinopathy (NPDR).³⁹ The FAZ-A is measured by segmenting the FAZ (demarcated as a blue region in Figure 1(c)) and calculating the total area using the following equation³⁵

$$\begin{array}{l}FAZ-A=[\mathit{Area}of\mathit{single}\mathit{pixel}\left(in\mu m^2\right)\\ \times {\displaystyle \sum }_{x=1,y=1}^{n}A(x,y)\hspace{0.17em}]\end{array}$$ (6)

where A(x, y) represents the pixels occupied by the segmented FAZ region.

FAZ-CI

FAZ-CI, also named as FAZ-circularity³⁸ or FAZ acircularity index,⁴⁶ measures the structural irregularity of the foveal shape.³⁸ FAZ-CI distortions have been observed in DR,⁴⁶ AMD,⁴⁷ SCR,³⁵ and glaucoma.⁴⁸ The FAZ-CI can be quantified by calculating the ratio of the perimeter of the FAZ-A to the perimeter of a reference circle with area identical to the FAZ³⁵

$$\mathit{FAZ}-CI=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n1}O(x,y)}{{{\displaystyle \sum }}_{x=1,y=1}^{n2}R(x,y)}$$ (7)

where O(x, y) represents the pixels occupied by the perimeter of the FAZ (green demarcation in Figure 1(c)), R(x, y) represents the pixels occupied by the perimeter of a reference circle with area identical to the segmented FAZ, n1 denotes the maximum number of pixels that encompasses the perimeter of the FAZ, n2 denotes the maximum number of pixels that encompasses the perimeter of the reference circle, x and y denote the pixel coordinates with respect to the perimeter of the FAZ or reference circle.

VCI

Fractal dimension (FD) has been used as a parameter to quantify the VCI (Figure 1(e)). FD is commonly calculated with the box-counting method in which a relationship is established by the number of boxes (of a certain resolution) that enclose the pattern in an image (Figure 1(e)).⁴⁹ This count is iteratively measured for different image scales. The box-counting method is formalized as⁴⁹

$$FD=\frac{\log \hspace{0.17em}{N}_r}{\log \hspace{0.17em}{r}^{-1}}$$ (8)

where $N_r$ is the number of boxes that encloses the pattern by the scale of the image, $r$. Lacunarity (LAC) is a complementary parameter to FD and is a measure of rotational inhomogeneity or the voids between vessel structures.⁵⁰ LAC provides information regarding the distribution and size of gaps in a binary image and is calculated using a similar box-counting strategy as the FD.

BPA

It is well known that decreased efficiency in blood transport can affect bifurcation in vascular structures.⁵¹ Quantitative BPA, including both angle and width features (Figure 2), has been recently demonstrated for objective OCTA classification of DR.⁵² For angle parameters, vessel branching angle and child (also referred to as daughter) branching angles (CBA1 and CBA2) are measured to quantify changes in bifurcation.⁵² As shown in Figure 2, in order to determine the branching angles, the branchpoint (green dots, Figure 2(b)) must first be determined, and then the vessel endpoints (red dots, Figure 2(c)). For the application in Figure 2, a radius of 0.097 mm was empirically determined and used to identify the vessel endpoints. It should be noted that there are no strict guidelines; however, the user should choose a radius that is large enough to contain the region of interest and small enough to not overlap with adjacent branchpoints. Furthermore, the application was for a 6 × 6 mm² FOV, therefore for other FOV, i.e. 3 × 3 mm² or 8 × 8 mm², the user should adjust accordingly. After identification of the vessel endpoints, using geometric identities the branchpoint angles, including the overall branching angle θ, and the individual child angles α1 and α2 (Figure 2(d)), are determined. Furthermore, an area of approximately 0.157 × 0.157 mm² surrounding the vessel endpoints was used to determine the width of each vessel.⁵² Similarly, the area size for width measurement should also be tailored to the specific application.

Figure 2.

(a) Sample branchpoint. (b) Branchpoint in a vessel skeleton, where the green pixel represents the branchpoint, the red pixels represent the end points, the blue pixels represent our vessels of interest, and the yellow circle represents the dilated area. (c) A composite image of the branchpoint (green) and endpoint (red), where the yellow square represents the window area. (D) Branch angle measurement. Angles A and B in the left image are complementary angles used to calculate θ, α1, and α2 in the right image. Source: Modified from Le et al.⁵² (A color version of this figure is available in the online journal.)

Utilizing the vessel widths, three width-based parameters were determined, i.e. the vessel branching coefficient and child width ratios are calculated to measure the structural change of the vessels as a result of bifurcation⁵²

$$\mathit{VBC}=\frac{d_1^2+d_2^2}{d_0^2}$$ (9)

$$CW{R}_i=\frac{d_i}{d_0}$$ (10)

where $d_0$ is the width of the parent vessel, $d_1$ and $d_2$ are the widths of the child vessels, and $i$ represents the $i$th child vessel. As noted by Le et al.,⁵² due to ischemia and neovascularization caused by DR, changes in vessel widths and branching angles could be related to vascular remodeling. We presume that branch geometry could be well correlated with other OCTA features such as BVT or VPI.

Differential A–V analysis

Differential A–V analysis compares the changes in arteries relative to the veins. Color fundus image analysis has been demonstrated to guide A–V classification in OCTA (Figure 3), showing improved sensitivity of OCTA detection of DR⁴¹ and SCR.^35,38 For color fundus image guidance, vessel nodes are identified as arteries and veins by evaluating optical density ratio between red and green channels of the fundus image.⁴¹ Once the source nodes have been identified, a fundus A–V map is generated. Employing image registration for the fundus and corresponding OCTA images, a vessel tracking algorithm is implemented to generate an A–V map in OCTA guided by the fundus A–V map.⁴¹ In addition to the color fundus image guided A–V differentiation in OCTA, OCT feature analysis guided A–V differentiation⁵³ and near infrared oximetry guided A–V differentiation⁵⁴ have been also demonstrated.

Figure 3.

A–V classification using fundus guided (row 1) and en-face OCT guided (row 2) techniques developed by Alam et al.^41,53 (a) Sample fundus image, (b) corresponding OCTA image, (c) OCTA A–V map overlaid on the fundus image, (d) sample OCT en-face image, (e) A–V information from en-face OCT overlaid on OCTA binary vessel map, and (f) OCTA A–V map. Source: Modified from Alam et al.^41,53 (A color version of this figure is available in the online journal.)

Flow analysis

Several parameters, including the flow index (FI),⁵⁵ adjusted flow index,²⁴ flow void (FV),⁵⁶ vascular connectivity,⁵⁷ and total retinal blood flow (TRBF), have been developed to quantify alterations in blood flow.⁵⁸ The information captured by OCTA is the decorrelation value of each pixel. Higher blood flow velocity results in increased decorrelation, i.e. enhanced OCTA brightness. FI is defined as the average decorrelation value within a region of interest in the en-face OCTA.⁵⁹ The FI can be measured by using the following equation

$$FI=\frac{{{\displaystyle \sum }}_{x=1,y=1}^{n}A(x,y)}{N}$$ (11)

where A(x, y) represents the decorrelation of the pixel (x, y) and N is total the number of pixels in the OCTA image.

The FV is calculated as the percentage of the area without flow signal over the total scanned region⁵⁶

$$FV=\frac{\mathit{Are}{a}_{\mathit{Flowvoid}}}{\mathit{Are}{a}_{\mathit{whole}}}\times 100\text{\%}$$ (12)

Because BVD represents the percentage of the area with flow signal over the total scanned region, the FV can also be determined by

$$FV=1-\mathit{BVD}$$ (13)

TRBF is a parameter used to estimate the flow of all vessel segments with Doppler OCT. TRBF is determined by integrating the axial velocity derived from the Doppler phase shift in the en-face plane⁵⁸

$$\mathit{TRBF}=-{\displaystyle {\iint }_{xy-\mathit{plane}}{v}_z\left(x,y\right)\text{d}x\text{d}y}$$ (14)

where the flow in a vessel is computed by integrating the axial flow velocity $v_z(x,y)$ measured from different pixels in the en-face image over the surface ($xy-\mathit{plane}$) normal to $v_z(x,y)$.⁵⁸ This method is primarily used in OCTA systems with high acquisition speeds, e.g. 100 kHz. For slower acquisition speeds, e.g. 70 kHz, strategies have been developed to determine TRBF using optimized en-face planes.^59,60

Another important OCTA parameter related to flow was described by variable interscan time analysis (VISTA),^61,62 which was performed to assess the alteration of choriocapillaris and differentiate varying degrees of flow impairment. VISTA has been used to quantify NPDR, PDR, geographic atrophy (GA), and AMD eyes.

CNV analysis

CNV analysis has been used to assess morphological distortions, named as seafan, medusa, tangled, and dead-tree, in choroidal vasculature.³⁰ A commonly examined parameter is CNV area, which can be determined as follows⁶³

$$\begin{array}{l}CNV\mathit{area}\left(m{m}^2\right)=\mathit{CNV}\mathit{area}\left(\mathit{pixel}\right)\\ \times {\left(3\hspace{0.17em}mm/304\mathit{pixel}\right)}^2\end{array}$$ (15)

where 3 mm represents the 3 mm FOV OCTA images (304 × 304 pixels) used in Uchida et al.⁶³

The essential step for quantitative OCTA analysis is reliable segmentation of CNV (Figure 4).

Figure 4.

Method of determining CNV area and skeleton. Illustrative steps of generating the CNV area and skeleton from the original outer retinal en-face OCTA image using a saliency model. Source: Reprinted from Patel et al.⁵⁷(A color version of this figure is available in the online journal.)

CNV: choroidal neovascular.

The segmentation has been commonly performed using manual or semi-manual processes.¹² Projection-resolved (PR-OCTA) method has been explored to determine the CNV area and skeleton automatically.⁵⁷ The binarized vessel or skeleton map of CNV area can be used for FD analysis using Fractalyse and likewise.³⁰ BVD or LAC can also be determined from the vessel or skeleton map. Other characteristics used to assess CNV include CNV location, CNV maturity, the presence of core vessels, and the presence of margin loops.⁶⁴ Quantitative OCTA parameters, such as BVD, FD,³⁰ and LAC,⁶⁵ have also been used for CNV analysis of AMD.³⁰ They are useful parameters to evaluate neovascularization activity by assessing the degree of complexity, and vessel nonuniformity of the lesions, respectively.

Discussion

Fundus photography has been established for clinical management of eye diseases. However, the spatial resolution and image contrast of traditional color fundus cameras are limited to reveal subtle distortion of retinal and choroidal vasculatures in early stages of eye diseases. OCTA provides a label-free solution for high resolution examination of ocular vasculatures. With depth-resolved capability to visualize retinal vasculatures at capillary level resolution, OCTA has been rapidly adopted for clinical management of eye diseases.

A brief summary of quantitative OCTA features established for clinical and custom OCTA instruments is listed in Table 1 along with demonstrated applications. Quantitative features, including BVT, BVC, BVD, VPI, FAZ-A, FAZ-CI, VCI, BPA, and differential A–V analysis, have been demonstrated to foster the standardization of objective interpretation of OCTA. Pathological mechanisms may affect the OCTA features sensitive to eye conditions. In other words, different eye diseases may cause OCTA distortions in different ways. For example, SCR is known to produce sickle shaped blood cells, which may lead to tortuous and dilated vessels that can be quantified by BVT and BVC, respectively. BVT has been recently demonstrated as the most sensitive feature for SCR detection and classification.²⁰ DR patients may frequently accompany with hypertension to cause arterial narrowing⁴¹ that can be quantified by BVC analysis. Capillary level ischemia can be assessed by BVD analysis in DR patients.³⁹ BVD has also been used to evaluate central and branch VOs.^{34,45,124,161,162} BVD, FAZ-A, and FAZ-CI have been commonly used for OCTA assessment of glaucoma.^48,78 Recent studies also suggest that vascular distortions may manifest in different regions due to different diseases. For example, localized BVD measurements revealed perifoveal region as the most sensitive region to classify NPDR stages.^39,79 However, for SCR, the parafoveal in the temporal retina was the most sensitive region for SCR staging.³⁸

Table 1.

Summary of quantitative OCTA features.

Features	Applications	Clinical OCTA with built-in software	Clinical OCTA with custom software	Custom OCTA
BVD	To quantify blood vessel density and identifies ischemic regions in DR, SCR, AMD, glaucoma, VO, etc.	21, 29, 47, 50, 66–77	22, 24, 25, 30 –33, 35–39, 41, 42, 45, 65, 72, 74, 78–112	113, 114
BVC	To quantify blood vessel width, dilation, or shrinkage in DR, SCR, etc.	115	22, 35, 37–41, 43, 85, 87–89, 95, 96, 99, 102, 108, 110, 116, 117	113, 114, 118
BVT	To quantify blood vessel tortuosity due to change in morphology in DR, VO, SCR, etc.	119	26, 33, 35, 38–41, 98, 101, 117, 120–123	114
VPI	To quantify changes in the perimeter of blood vessels in DR, VO, SCR, etc.	–	22, 35, 38, 39, 42, 95, 96, 102, 110, 112	–
FAZ-A	To quantify changes in foveal area in DR, DME, SCR, etc.	21, 26–29, 34, 44, 47, 50, 66–71,73, 75, 77, 115, 119, 124–137	24, 26, 33, 35, 38, 39, 43–46, 48, 72, 74, 78, 79, 81, 89–94, 98, 100, 101, 103, 105, 107, 108, 110, 119, 132, 138–148	–
FAZ-CI	To quantify changes in foveal complexity in DR, DME, VO, AMD and SCR.	47, 48, 66, 67, 70, 72, 136	46, 78, 79, 89, 90, 92, 110, 132, 139, 142 –144	–
VCI	To quantify changes in retinal vascular complexity in DR, SCR, etc.	–	30, 32, 36, 37, 49, 65, 79–88, 90, 102, 104, 149, 150–154	–
BPA	To quantify changes in vascular bifurcation in DR, etc.	–	52	–
A–V analysis	To achieve differential artery and vein analysis in DR, SCR, etc.	–	40, 41, 116	54
Flow analysis	To quantify changes in blood flow in DR, GA, AMD, etc.	21, 125	24, 55–57, 106, 109, 138, 155, 156	15, 61, 157
CNV analysis	To quantify neovascularization activity in AMD, etc.	–	30, 57, 63–65, 82, 117, 149, 151, 156, 158–160	15, 61, 159

AMD: age-related macular degeneration; A–V: artery–vein; AVR: artery–vein ratio; BPA: branchpoint analysis; BVC: blood vessel caliber; BVD: blood vessel density; BVT: blood vessel tortuosity; CI: contour irregularity; CNV: choroidal neovascular; DME: diabetic macular edema; DR: diabetic retinopathy; FAZ: foveal avascular zone; FAZ-A: fovea avascular zone area; FAZ-CI: FAZ contour irregularity; FD: fractal dimension; GA: geographic atrophy; LAC: lacunarity; OCTA: OCT angiography; SCR: sickle cell retinopathy; VCI: vessel complexity index; VO: vein occlusion; VPI: vessel perimeter index.

Differential A–V analysis has been demonstrated to improve OCTA performance.^40,41 Pathological alterations in the artery and vein compartments are known to be affected in different ways.¹⁶³ For example, DR may cause increased arterial tortuosity, venous beading, narrowing artery, and dilated vein.¹⁶⁴ If only mean value of arterial and venous distortions is evaluated between control and diseased eye, the OCTA sensitivity can be compromised. A recent study revealed two differential A–V features, i.e. artery vein ratio of blood vessel caliber and tortuosity to improve the OCTA performance for DR classification,⁴¹ and predominant BVD and BVT distortions were observed in venous system in case of SCR.⁴⁰

Quantitative analysis of choroidal vasculature has been explored predominantly for AMD assessment.^61,158 Dry AMD is generally characterized by decreased foveal choroidal blood flow and increased drusen and GA. In case of wet AMD, CNVs are the most commonly used biomarkers for quantitative OCTA analysis.^61,165 Jia et al.¹⁵ first presented a study focused on using OCTA to detect and classify the types of CNVs. This study reported that the en-face angiograms showed decreased choroidal flow adjacent to the CNV in all cases. Cross-sectional angiograms were able to visualize location and classify the CNV as type I and type II. In following studies,⁵⁷ the researchers analyzed artifact removal algorithms, slab subtraction, and PR-OCTA for quantification of CNV area and connectivity. Miller et al.¹⁵⁹ compared CNV using SS and SD-OCTA, and reported statistically significant differences between SS-OCTA and SD-OCTA measured lesion areas. SS-OCTA showed larger lesion area when compared to SD-OCTA in both 3 × 3 mm² and 6 × 6 mm², with larger differences reported in the 6 × 6 mm². Similarly, Zhang et al.⁵⁶ reported that SS-OCTA allowed for deeper light penetration into the choroid. Another parameter that has been studied for AMD is FD. Al-Sheikh et al.³⁰ studied CNV lesions in AMD pre- and post-treatment of anti-vascular endothelial growth factors, and reported a lower FD in the inner part of the lesion after treatment.

Quantitative OCTA opens a unique opportunity to enable computer-aided disease detection and AI classification of different eye diseases. Machine learning techniques have been explored to segment microvasculature,¹⁶⁶ non-perfusion areas,¹⁶⁷ optical nerve head.¹⁶⁸ A support vector machine (SVM) classifier was recently demonstrated for supervised machine learning based OCTA classification of SCR³⁵ and DR.³⁹ Six OCTA features, i.e. BVD, BVC, BVT, VPI, FAZ-A, and FAZ-CI, were used to train the SVM classifier for identifying control, mild, and severe SCR subjects.³⁵ The SVM was able to identify control versus disease and mild versus severe SCR with 100 and 97% accuracies, respectively. A similar approach was also adopted for OCTA classification of NPDR.³⁹ The study demonstrated 94.41% accuracy for control versus disease (i.e. NPDR) and 92.96% accuracy for control versus mild NPDR classification. The control versus mild NPDR classification was important for early DR detection. In another study, three OCTA features, i.e. BVD, BVC, and FAZ-A, were used for automated classification of NPDR with 94.3% accuracy.¹⁰⁸ Supervised machine learning was also recently validated for multiple-task classification to differentiate control, SCR and DR from each other.²⁰ In principle, both supervised and unsupervised machine learning techniques can be considered for AI classification of OCTA images. However, the limited size of currently available database is a major challenging factor for deep learning-based AI classification of OCTA. With the increasing OCTA applications in ophthalmology to expand the available OCTA datasets, we anticipate that deep learning-based classification, which has been well established for fundus image classification, will play an important role to enable automated OCTA detection and classification of eye conditions. Alternatively, transfer learning-based technology¹⁶⁹ may find valuable application in the near future to foster the deep learning OCTA classification of eye diseases.

Authors’ contributions

Original draft preparation: XY, MNA, DL; manuscript review and editing: XY and DT; project supervision: XY.

DECLARATION OF CONFLICTING INTERESTS

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

This research was supported in part by NIH grants R01 EY030101, R01EY029673, R01 EY023522, P30 EY001792; by unrestricted grant from Research to Prevent Blindness; by Richard and Loan Hill endowment.

ORCID iDs

Minhaj N Alam https://orcid.org/0000-0003-3095-2232

David Le https://orcid.org/0000-0003-3772-1875