Optical Coherence Tomography Angiography of Volumetric Arteriovenous Relationships in the Healthy Macula and Their Derangement in Disease

Abstract

Purpose: To characterize relative arteriovenous connectivity of the healthy macula imaged by optical coherence tomography angiography (OCTA) using a new volumetric tool.

Methods: OCTA volumes were obtained for 20 healthy controls (20 eyes). Two graders identified superficial arterioles and venules. We implemented a custom watershed algorithm to identify capillaries most closely connected to arterioles and venules by using the large vessels as seeds to flood the vascular network. We calculated ratios of arteriolar- to venular-connected capillaries (A/V ratios) and adjusted flow indices (AFIs) for superficial capillary plexuses (SCPs), middle capillary plexuses (MCPs), and deep capillary plexuses (DCPs). We also analyzed two eyes with proliferative diabetic retinopathy (PDR) and one eye with macular telangiectasia (MacTel) to evaluate the utility of this method in visualizing pathological vascular connectivity.

Results: In healthy eyes, the MCP showed a greater proportion of arteriolar-connected vessels than the SCP and DCP (all P < 0.001). In the SCP, the arteriolar-connected AFI exceeded the venular-connected AFI, but this pattern reversed in the MCP and DCP, with higher venular-connected AFI (all P < 0.001). In PDR eyes, preretinal neovascularization originated from venules, whereas intraretinal microvascular abnormalities were heterogeneous, with some originating from venules and others representing dilated MCP capillary loops. In MacTel, diving SCP venules formed the epicenter of the outer retinal anomalous vascular network.

Conclusions: Healthy eyes showed a higher MCP A/V ratio but relatively slower arteriolar vs. venular flow velocity in the MCP and DCP, which may explain deep retinal vulnerability to ischemia. In eyes with complex vascular pathology, our connectivity findings were consistent with histopathologic studies.

The retina is one of the most metabolically active tissues in the body and receives a dual blood supply from the retinal capillaries and the choriocapillaris.¹ In the parafovea, the retinal capillaries are comprised of three layers—the superficial capillary plexus (SCP), middle capillary plexus (MCP), and deep capillary plexus (DCP).^2–6 Although the general structure of the parafoveal retinal capillary system is well established, the communication among these layers is complex and remains a subject of debate. Several different models have been proposed to describe arteriovenous connectivity between these layers. One model is based on microscopy studies in pigs and mice showing sequential flow from the SCP to the MCP to the DCP. This model proposes a primarily serial organization of the retinal layers, with venous drainage from the DCP.^7–9 Another model suggests that each capillary bed has a “hammock” arrangement and is based on histological studies that show distinct arteriolar supply and venous drainage from each layer.^4,10 Finally, a hybrid model has also been proposed that combines features of both models, with independent supply and drainage of each bed coupled with numerous interconnections and anastomoses between the layers.^11,12 

Retinal vascular diseases may differentially affect arterioles and venules. For example, in diabetic retinopathy (DR), dilation, tortuosity, and eventually neovascularization primarily occur on the venous side.^13,14 This is supported by primate studies where intravitreal vascular endothelial growth factor–injected eyes developed neovascularization from venules.¹⁵ Another example occurs in macular telangiectasia type 2 (MacTel), a rare disease thought to be of neurodegenerative origin, where right-angle venules connect directly to anomalous capillaries in the deep retina.¹⁶ 

Optical coherence tomography angiography (OCTA) is a non-invasive in vivo imaging modality that uses repeated structural scans to extract retinal vessel blood flow and structure.¹⁷ A small number of studies have explored differential arteriovenous involvement in retinal disease using OCTA. Ishibazawa et al.,¹⁸ using widefield OCTA in eyes with DR, suggested that nonperfusion in DR predominantly localized around arterioles. More recently, our group found that decreased vessel density was more prominent surrounding venules than arterioles in early MacTel.¹⁹ However, several limitations of these studies must be considered. Notably, they defined periarteriolar and perivenular regions by proximity to nearest arteriole or venule. These approaches do not account for diving or overlapping vessel connections to the deeper layers, precluding meaningful evaluation of arteriovenous relationships in the MCP and DCP. They also may double-count capillaries in regions where arterioles and venules overlap. 

In this study, we introduced a novel volumetric method to identify arteriolar- and venular-connected capillaries from three-dimensional (3D) OCTA slabs, which we use to characterize arteriovenous differences in healthy eyes. We also explored clinical applications of this method in eyes with retinal vascular pathology. 

Materials and Methods

This retrospective study analyzed healthy patients imaged by OCTA between June 2015 and March 2019 at the Department of Ophthalmology at Northwestern University in Chicago, Illinois. The study was approved by the Institutional Review Board of Northwestern University and conducted in accordance with the tenets of the Declaration of Helsinki and the regulations of the Health Insurance Portability and Accountability Act. Written informed consent was obtained from all subjects. We included healthy controls who were 20 to 30 years old with no history of ocular pathology, intraocular surgery, or systemic disease such as hypertension or diabetes. We excluded eyes with refractive error greater than 6.0 diopters (D), astigmatism greater than 3.0 D, and significant media or lens opacities. We also imaged three eyes of patients with retinal vascular pathologies as part of a separate qualitative analysis. 

OCTA Imaging

OCTA images were obtained using the RTVue-XR Avanti system (Optovue, Inc., Fremont, CA, USA) with the split-spectrum amplitude-decorrelation angiography (SSADA) algorithm (version 2017.1.0.151).¹⁷ This system captures two consecutive B-scans, each containing 304 A-scans, in the approximately 3 mm × 3 mm region centered on the fovea, with an A-scan rate of 70,000 scans/s using a light source centered on 840 nm and a bandwidth of 45 nm. For quantitative analysis of healthy controls, we excluded images with a quality index (Q-score) less than 9 or signal strength index (SSI) less than 70. One eye per patient was analyzed. If both eyes of a patient met the quality criteria, we selected the eye with the higher SSI. 

We segmented the SCP, MCP, and DCP as previously described,²⁰ with the SCP obtained from the internal limiting membrane (ILM) to 10 µm above the inner plexiform layer (IPL), the MCP from 10 µm above to 30 µm below the IPL, and the DCP from 30 µm below the IPL to 10 µm below the outer plexiform layer. We exported two-dimensional (2D) angiographic images of each layer with projection artifacts removed using the built-in Optovue software. We also exported raw 3D volumes of each scan and removed projection artifacts using a custom MATLAB (MathWorks, Natick, MA, USA) program based on the method described by Zhang et al.²¹ 

Image Analysis

All image analysis was performed using FIJI, an open-source distribution of the program ImageJ (National Institutes of Health, Bethesda, MD, USA).²² Our proposed algorithm to identify arteriolar and venular capillary connectivity was based on watershed segmentation, using the arterioles and venules of the SCP as start points from which to flood the remaining capillary networks in 3D (Fig. 1).²³ We then postprocessed the resulting 3D stacks to generate 3D volumes and 2D images of the SCP, MCP, and DCP, with vessels labeled by their connectivity to the nearest arteriole or venule. 

Figure 1.

 

Image processing workflow for assigning arteriolar and venular connectivity. Iterative watershed segmentations were performed to gradually flood the 3D vascular slab starting from the labeled SCP large vessels. Postprocessing was then performed to generate 2D images of arteriolar and venular connectivity from which OCTA parameters were calculated. Note that OCTA parameters were calculated from grayscale images prior to recoloring.

Generation of Large Vessel Seed Volumes

We first sought to generate 3D volumes containing only the SCP large vessels, labeled by color to distinguish arterioles and venules. Arterioles and venules were first identified by tracing vessel paths from the corresponding scanning laser ophthalmoscopy fundus image obtained by the OCTA device and confirmed in the OCTA scan based on features including capillary clearing directly surrounding arterioles.¹² To isolate the large vessel shapes, we binarized the 2D SCP slab using MaxEntropy thresholding (Fig. 2B), followed by applying a circularity cutoff of 0.40 to exclude speckle noise. We then reconnected large vessels using morphologic dilation followed by rebinarization with MaxEntropy thresholding (Fig. 2C). The 2D large vessel maps were manually corrected to remove any remaining shapes that could not be confirmed to be attached to large vessels and to add in missed vessels. Arterioles and venules were color-coded using white for arterioles and gray for venules (Fig. 2D). 

Figure 2.

 

Generation of labeled SCP large vessel maps. (A) SCP large vessels were identified as arterioles (red) or venules (blue) based on the presence of capillary clearing surrounding arterioles and the draining of venules into deep vortices. (B) SCP images were first binarized with MaxEntropy thresholding. (C) Speckle noise was removed by applying a circularity cutoff and vessels that had become fragmented by despeckling were reconnected through morphologic operations. (D) Binarized maps manually corrected as necessary and recolored to label arterioles (white) and venules (gray).

The large vessels were isolated from the 3D SCP volume by masking their corresponding 2D outlines, followed by MaxEntropy thresholding to generate a 3D volume containing only binarized large vessels. We color-coded arterioles and venules in the 3D volume with the same scheme as above. Areas where arterioles and venules overlapped were manually recolored by referencing the 3D SCP volume to ensure accurate determination of whether the arteriole was on top of the venule or vice versa. To evaluate the intergrader reproducibility of the arteriole and venule 3D volumes, a second masked grader performed the same operations in a random subset of 10 eyes. 

Watershed Segmentation

The 3D large vessel volumes were then used as markers for subsequent watershed segmentation using the MorphoLibJ plugin in FIJI.²⁴ Normally, watershed segmentation uses a labeled marker image that represents the starting points for the segmentation. From the markers, the algorithm then floods a binary mask image, assigning pixels in the mask based on how close they are to the starting marker points.²³ However, distinguishing between vessel signal and background noise in OCTA images is imperfect, as projection and decorrelation artifacts can introduce speckling and apparent breaks in signal within vessels.²⁵ Binarization that uses a high (strict) threshold identifies connections with high certainty but may fail to identify vessels that are interrupted by artifact or shadowed by overlying vessels. Conversely, using a low (more lenient) threshold will connect more vessels but also include more noise pixels. 

To address the tradeoff between certainty of vessel identification and total percentage of vessels assigned, we implemented an iterative thresholding algorithm for watershed segmentation. This method starts with a 3D volume of the retinal vessels binarized using a strict (high) threshold as the mask and uses the arterioles and venules (Fig. 1, large SCP vessels) as the starting points for the watershed algorithm. This watershed algorithm categorizes the vessels in the 3D volume as arteriolar, venular, or unconnected (Fig. 1, watershed). The threshold used to binarize the vessel network is then gradually lowered to reveal more lower-intensity vessels. With each iteration, we repeat the watershed algorithm, using the previous categorized vessel image (watershed) to flood the newly exposed vessels. As a result, vessels that can be identified as arteriolar- or venular-connected with high certainty at higher thresholds are preserved, and iterating through lower threshold values reduces the fraction of unconnected vessels (Fig. 3). 

Figure 3.

 

Iterative thresholding mitigates the tradeoff between accuracy and percentage of vessel assignment as threshold decreases. Rows indicate the SCP (top), MCP (middle), and DCP (bottom), respectively. Columns compare vessel assignments using only high (left), low (middle), and iterative (right) thresholding methods. At high thresholds (A, D, G), few vessels are identified but there is a high certainty of underlying connectivity (yellow dashed boxes). Low thresholds (B, E, H) assign many vessels but may make inaccurate connections (cyan dashed boxes). Iterative thresholding (C, F, I) preserves more accurate vessel assignments made at higher thresholds, but also achieves the higher percentage of vessel assignment seen at lower thresholds.

Postprocessing

We used maximum projection on the 3D volumes to generate 2D masks of arteriolar and venular vessel classifications. These masks were used to identify arteriolar- and venular-connected vessels from the original OCTA 2D scans. To create color renderings of arteriolar and venular connectivity, we overlaid the arteriolar and venular images, color-coding the arteriolar-connected vessels in red and venular-connected vessels in blue. 

Data Analysis

To assess the accuracy of venous connectivity in the deeper retinal layers, we evaluated the accuracy of venous vortex identification in the DCP.^11,26 We sought to compare our proposed watershed algorithm with radius-based methods, which define arteriolar and venular capillaries as those occurring within a certain radius of SCP arterioles and venules.^18,19 Specifically, for the radius-based method, we used the same SCP arteriole and venule outlines but defined arteriolar and venular capillaries as those occurring within a 0.15-mm radius of an arteriole or venule, respectively.¹⁹ Venous vortices were identified from the DCP slabs and compared to the arteriovenous assignments generated by each method. We quantified the percentage of vortices that were correctly identified as being of venous origin (blue). 

Analysis of quantitative OCTA parameters was performed using the original grayscale image outputs prior to any recoloring. We determined the noise threshold for images using mean thresholding. Arteriovenous (A/V) vessel ratios were calculated from the 3D volumes as the ratio of arteriolar-connected vessels to venular-connected vessels for each layer.¹⁹ We calculated an SCP capillary-only A/V ratio by excluding the large vessels from consideration as previously described.²⁷ We determined length A/V ratios for each layer from 3D volumes by skeletonizing vessels to a single pixel width. We also analyzed A/V ratios for the 2D OCTA scans as the ratio of area occupied by arteriolar- vs. venular-connected vessels. Each adjusted flow index (AFI), a proxy for flow velocity, was calculated using 2D OCTA scans for arteriolar- and venular-connected vessels in each layer as the average decorrelation value of all pixels above the noise threshold.^28,29 

Statistical Analysis

Statistical analysis was performed using SPSS Statistics 26 (IBM Corporation, Chicago, IL, USA). Shapiro–Wilk tests showed that A/V ratio and AFI data were normally distributed. We compared A/V ratios between the capillary plexuses within each eye using paired t-tests. P-values were adjusted for multiple comparisons using the Benjamini–Hochberg method with the false-discovery rate set at 0.05 to minimize both type I and type II errors.³⁰ We also compared arteriolar- versus venular-connected AFIs using paired t-tests. We calculated intraclass correlation coefficients (ICCs) to determine the interrater reliability of arteriole and venule grading. For eyes that had repeat OCTA imaging within a week of baseline imaging, we also calculated ICCs to assess interscan reliability. 

Results

Twenty eyes of 20 healthy subjects were analyzed. Mean age of the subjects was 25.3 ± 2.2 years (range, 22–30); 10 of the subjects (50%) were female. All eyes had a Q-score of 9, with average SSI of 76.9 ± 3.8. There were no significant differences between male and female subjects with respect to age (P = 0.486) or image SSI (P = 0.776). Overall, the watershed segmentation method with iterative thresholding achieved excellent assignment of vessels in the SCP (99.61% ± 0.48%) and MCP (85.14% ± 3.04%) but lower overall rate of assigned capillaries in the DCP (59.12% ± 10.03%). Interrater ICCs for generation of the large vessel seed volumes ranged from 0.893 to 0.940 for calculated A/V ratios (all P < 0.001) and 0.996 to 1.000 for AFIs (all P < 0.001), showing good to excellent reliability for these parameters. Twelve eyes had repeat imaging, and interscan ICCs ranged from 0.719 to 0.851 for calculated A/V ratios (all P ≤ 0.003), showing good repeatability. However, interscan ICCs for AFIs had good repeatability in the SCP and MCP (0.804–0.926; all P < 0.001), but not in the DCP, with 0.299 (P = 0.160) and 0.553 (P = 0.025) for arteriolar-connected and venular-connected AFIs, respectively. A similar pattern was seen for overall scan AFIs, with ICCs of 0.816 to 0.905 in the SCP and MCP (all P < 0.001), but 0.426 (P = 0.073) in the DCP, suggesting that the lack of DCP AFI repeatability is inherent to inter-image variation rather than errors by the proposed algorithm. 

On average, our algorithm identified DCP venous vortices with 90.6% ± 7.8% accuracy, which was significantly improved from the 68.1% ± 16.6% achieved by the radius-based method (P < 0.001) (Fig. 4). The radius-based method tended to incorrectly classify DCP vortices as being of both arteriolar and venular origin (Fig. 4C, cyan solid circles) or divided vortices between arteriolar and venular regions (Fig. 4C, white asterisk). Incorrect identification of DCP vortices using our proposed method resulted when vortices were too close to the edge of the image for their draining venules to be captured in the frame or were due to the presence of substantial SCP vessel shadow artifact (Fig. 4D, white arrow) that obscured underlying vascular connections. 

Figure 4.

 

Assessing accuracy of the algorithm in assigning venous vortices in the deep capillary plexus. (A) SCP with color overlay indicating arterioles (red) and venules (blue). (B) DCP with vortex veins identified (dashed circles) and overlaid SCP arterioles (red) and venules (blue). (C) DCP capillaries classified using the radius-based method. Capillaries were classified as arteriolar (red) or venular (blue) if they were within a 0.15-mm radius of SCP arterioles or venules, respectively. Unassigned capillaries are in gray. Magenta vessels represent capillaries that fell within 0.15 mm of both an arteriole and a venule. This method incorrectly identified many DCP vortices (cyan solid circles) as being of both arteriolar and venular origin or divided vortices between multiple regions (white asterisk). (D) DCP capillaries classified using our proposed watershed algorithm method, with arteriolar-connected in red and venular-connected in blue. Incorrectly identified vortices using this method typically resulted from large superficial vessel shadows (white arrow).

When evaluating OCTA parameters, we considered the SCP capillaries separately from the overall SCP slab to eliminate the effects of the large vessels. Overall, we found that the MCP showed a greater proportion of arteriolar-connected vessels than both the SCP capillaries and the DCP, regardless of whether the A/V ratios were calculated by vessel volume (all P < 0.001) or vessel length (all P < 0.002) (Fig. 5). Area A/V ratios calculated from 2D images also showed a similar pattern of a higher proportion of arteriolar-connected vessels in the MCP compared to the SCP capillaries and DCP (all P < 0.001; Supplementary Fig. S1). 

Figure 5.

 

A/V ratios show consistent patterns of variation among the SCP, MCP, and DCP of the macula in healthy eyes. (A) A/V ratios calculated by volume. (B) A/V ratios calculated by length. Overall, A/V ratios were lowest in the SCP capillaries, indicating a larger proportion of venular-connected vessels, and highest in the MCP, indicating a larger proportion of arteriolar-connected vessels. *P < 0.05, **P < 0.01.

Figure 6.

 

Relative relationships between AFIs of arteriolar- and venular-connected vessels in the different capillary layers. Arteriolar-connected vessels are indicated in red, venular-connected vessels in blue, and unassigned vessels in gray. In the SCP capillaries (A, E) and whole SCP (B, F), AFIs of arteriolar-connected vessels are significantly higher than those of venular-connected vessels (all P < 0.001). This pattern reverses in the MCP (C, G) and DCP (D, H), where AFI of venular-connected vessels is significantly greater (all P < 0.001). Unassigned vessels were not included in AFI calculations. *P < 0.05, **P < 0.01.

We further explored the relationships between AFIs in arteriolar-connected versus venular-connected vessels within each layer (Fig. 6). We found that arteriolar-connected vessels in the SCP capillaries and whole SCPs had significantly higher AFIs than their venular-connected counterparts (both P < 0.001). However, this pattern was reversed in the MCP and DCP, with significantly higher venular-connected compared to arteriolar-connected AFIs (both P < 0.001). We were unable to compare AFI between layers due to the difference in total percentage of assigned vessels in each layer. 

We then used this algorithm to qualitatively analyze eyes with retinal vascular pathology. We imaged both eyes of a 42-year-old male with proliferative diabetic retinopathy (PDR) who had no history of ocular surgery or treatment in either eye. One of the PDR eyes was noted to have preretinal neovascularization bordering the fovea, and the other had multiple intraretinal microvascular abnormalities (IRMAs). We also analyzed an eye of a 56-year-old male with outer retinal vascular abnormalities due to MacTel that had not been previously treated. All pathologic eyes had Q-scores ≥ 8 and SSIs ≥ 68. 

In the PDR eyes, the algorithm successfully identified the neovascular frond (Fig. 7B; yellow arrows) as originating from the venous side (Fig. 7D). In contrast, IRMA lesions were more heterogeneous, with some appearing as venular branches (Fig. 8G, cyan and magenta dotted circles) and others as loops of dilated capillary connecting an arteriole and venule within the MCP (Figs. 8G, 8H, yellow dotted circles). Notably, in the PDR eye with neovascularization, few vessels outside the perifoveal region were successfully assigned in the MCP and DCP, likely a result of the overall low vessel density due to advanced ischemia in PDR, a potential limitation of the algorithm. 

Figure 7.

 

Arteriolar and venular connectivity in an eye with PDR and neovascularization (yellow arrows) in the macula. Note the predominantly venous connectivity of the neovascular frond (D, yellow dotted circle). The substantial capillary dropout seen in PDR also results in more unassigned vessels in the MCP (E) and DCP (F). (A) SCP with arterioles in red and venules in blue. (B) B-scan corresponding to white dashed line in A with the flow overlay in red and neovascularization (yellow arrow) extending into the preretinal space. (C) Labeled SCP large vessel map used for watershed algorithm. (D–F) Postprocessed images showing connectivity in the SCP (D), MCP (E), and DCP (F). Arteriolar-connected vessels are in red, venular-connected vessels in blue, and unassigned vessels in gray.

Figure 8.

 

Arteriolar and venular connectivity in an eye with PDR and IRMAs (arrows) in the macula. Three areas of IRMAs (cyan, magenta, and yellow arrows) were identified. Two IRMAs (cyan and magenta dotted circles) were labeled as venular-connected, and one IRMA appeared as a dilated capillary loop (yellow dotted circle) interconnecting an arteriole and venule. (A) SCP with arterioles in red and venules in blue. (B) Full retinal slab showing the vascular network surrounding each IRMA. (C) Labeled SCP large vessel map used for watershed algorithm. (D–F) B-scans corresponding to each IRMA labeled in A with flow overlay in red. Note that the IRMA is located below the ILM, unlike the preretinal neovascularization in Figure 7, which extends above the ILM. (G–I) Postprocessed images showing connectivity in the SCP (G), MCP (H), and DCP (I). Arteriolar-connected vessels are in red, venular-connected vessels in blue, and unassigned vessels in gray.

In the MacTel eye, we visualized two separate areas where diving venules originating in the SCP converged into an area of anomalous outer retinal vasculature (Figs. 9E–9H, cyan and yellow dotted circles). In one of these lesions, a superior venule appeared to cross the horizontal raphe to connect to an inferior venule (Figs. 9E–9H, yellow dotted circles). The arteriovenous distribution of the outer retinal anomaly (Fig. 9H) is congruous with the general vascular distribution of the DCP layer (Fig. 9G). The foveal avascular zone is completely obliterated by these outer retinal vascular anomalies. 

Figure 9.

 

Arteriolar and venular connectivity in an eye with outer retinal vascular abnormality due to MacTel. The area of outer retinal vascular abnormality connects to diving venules that converge toward two central points (cyan and yellow dotted circles), suggesting that outer retinal vascular changes in MacTel may reflect deformity of the overlying DCP. (A) SCP with arterioles in red and venules in blue. (B) Labeled large vessel map used for watershed algorithm. (C) B-scan corresponding to white dashed line in A with flow overlay in red. Note the vessels extending into the outer retinal layer (yellow arrow). (D) En face scan of the outer retinal vascular abnormality segmented as shown in C (green lines). (E–H) Postprocessed images showing arteriolar and venular connectivity in the SCP (E), MCP (F), DCP (G), and the outer retinal vascular abnormality (H).

Discussion

In this study, we developed a method for assigning 3D, volumetric arteriovenous connectivity of the macular capillary plexuses on OCTA. This method achieved excellent accuracy in identifying DCP venous vortices and showed good intergrader reliability, supporting its utility. We showed that, in healthy eyes, the MCP is characterized by higher proportion of arteriolar connectivity but relatively slower flow in arteriolar-connected vessels. We also demonstrated that, in eyes with retinal vascular pathology, our findings were consistent with histopathologic studies. 

Previous studies of retinal vascular connectivity have relied on tracing individual capillary paths from arteriole to venule to establish connections between the different retinal plexuses.^11,12,31,32 However, this approach has typically been confined to limited areas on histologic preparations, requires significant manual input, and does not capture the bulk contribution of these connections to the overall arteriolar and venular supply and drainage of each plexus. Furthermore, these postmortem histopathologic techniques are subject to tissue artifacts and are not suitable for estimating blood flow. Other studies have attempted to define arteriolar and venular capillaries based on radial distance from SCP arterioles and venules.^18,19 This method also has several limitations. It does not accurately classify capillaries when large vessels cross and is unreliable in the deeper layers, where diving large vessel branches may result in a different distribution of arteriolar and venular capillaries that does not mirror the overlying SCP. Strengths of our current method include the watershed algorithm, which in effect “traces” multiple paths at once and is based on direct connectivity rather than general proximity of capillaries to arterioles or venules, allowing analysis of the deeper capillary layers, and the iterative approach to prioritize connections of higher certainty. The use of OCTA imaging provides in vivo visualization of vascular flow rather than structure, which makes it more applicable to clinical settings. 

Our finding of higher arteriolar connection relative to venous in the MCP of healthy eyes is consistent with recent data from human histologic studies,^11,31 suggesting that our method provides a non-invasive alternative for accurately characterizing and quantifying arteriovenous connectivity in living eyes. In donor eyes, An et al.¹¹ showed that the MCP capillaries are supplied from MCP-direct arteriolar branches as well as from indirect arteriolar branches from the SCP, whereas the DCP only receives arteriolar supply indirectly via the MCP. Notably, Cabral et al.,³² in an OCTA-based study, also suggested that the shortest path for most capillary connections between arterioles and venules passed through the MCP, with few to no direct SCP-to-SCP connections. The dual supply of the MCP from the SCP and MCP arteriolar branches, as well as the routing of arteriolar shortest paths through this layer, could contribute to the predominance of arteriolar-connected capillaries relative to venous that we observed in the MCP. 

The existence of venous vortices in the DCP has been well documented in a variety of histologic^11,33 and OCTA studies.^26,32,34 However, their exact connectivity to large superficial venules and significance in the macular circulation are still debated. Some investigators have suggested that the DCP serves as a primarily venous drainage route from the SCP and MCP,^8,34 whereas others suggest that the DCP receives an independent arteriolar supply and venous drainage.¹² Although our current study, as a bulk analysis of arteriovenous connectivity, was not designed to visualize individual capillary paths through each layer, we did show an apparent predominance of venular-connected capillaries in the DCP relative to the MCP. All three capillary layers are thought to drain directly into the large SCP venules.¹¹ However, connections from the SCP and MCP are primarily comprised of two capillaries merging into a venule. Venous vortices composed of multiple converging capillaries are only seen in the DCP, which may explain the greater proportion of venular-connected capillaries we observed in that layer.¹¹ A lack of direct arteriolar channels from the SCP to the DCP could also contribute to the relatively lower proportion of arteriolar-connected vessels in the DCP.^11,31 A final consideration includes the contribution of image artifacts to capillary density measurements, as the physiologic capillary-free zone surrounding arterioles in the SCP may be artificially propagated into the MCP and DCP on certain OCTA systems.³⁵ However, this phenomenon would be expected to decrease relative arteriolar connectivity and thus A/V ratio in all layers, not just the DCP, and is unlikely to have substantially affected our overall conclusions. 

We found that arteriolar-side flow velocities exceeded venular-side flow velocities in the SCP, consistent with the smaller diameter of arterioles compared to venules.³⁶ However, this pattern unexpectedly shifted in the MCP and DCP, where venular-side flow velocities exceeded arteriolar. Although we do not have a definitive explanation for these findings, we can consider several possibilities. Conservation of mass would require that the volumetric rates of blood flow entering and exiting each capillary bed are equal. The AFI indirectly measures fluid velocity,³⁷ which depends on the total cross-sectional area of vessels. Division of arteriolar flow among more capillaries, as indicated by the higher A/V ratio in the MCP, would thus be expected to increase total cross-sectional area and decrease arteriolar-side flow velocity relative to the venular side. 

However, arteriolar-side AFI remained lower than venular in the DCP, which conversely had more venular-connected vessels than the MCP, suggesting the relationship of AFI with A/V ratio may be more complex than a simple inverse correlation. For example, flow velocities in each capillary bed could also be influenced by presence of anastomotic vessels between layers,^11,31 which are vertically oriented and therefore could not be visualized on OCTA imaging. The MCP and DCP in particular have been shown in rodent studies to be highly interconnected by such anastomoses.³⁸ We acknowledge that our interpretation of these AFI findings is speculative, as AFI does not directly measure blood flow and correlates to flow velocity only within a limited range.³⁷ Directly comparing AFIs among layers was also not possible in this study because of the non-uniform percentage of unassigned vessels in each layer; more unassigned vessels in the deeper layers elevated the mean AFIs in those layers. AFI is also dependent on imaging depth, which further precludes comparisons among layers.³⁷ Interpretation of these results is also limited by the overall low interscan repeatability of AFI measurements in the DCP, and further studies comparing arteriovenous OCTA flow findings to tools that directly measure flow such as Doppler or adaptive optics scanning laser ophthalmoscopy imaging are needed. 

These differences in arteriovenous predominance and flow among the three layers have interesting potential implications for our understanding of retinal vascular function and disease. Notably, the deep retinal layers are thought to be more vulnerable to damage in certain retinal diseases. For example, hemorrhages and microaneurysms in DR develop earlier in the deep capillaries,^39,40 and longitudinal studies have shown that deep capillary nonperfusion predicts a variety of DR complications.^41–43 Paracentral acute middle maculopathy and acute macular neuroretinopathy, which result in visual scotomas and are characterized by outer retinal lesions on OCT imaging, are thought to be associated with vascular compromise at the MCP and DCP.^44–46 It is possible that the MCP may represent a transition zone where more arteriolar capillary branches results in relative slowing of arteriolar-side flow velocity into the MCP and secondarily the DCP. These hemodynamic characteristics of the MCP and DCP may contribute to their apparent vulnerability to ischemia. 

Vascular abnormalities including IRMAs and neovascularization are key features of DR. In PDR, histologic studies have suggested that preretinal neovascularization originates from venules.^14,47 However, IRMAs may appear as dilated capillary segments or abnormal vessels arising from arterioles or venules.^48,49 Although the presence of IRMAs is associated with subsequent development of neovascularization,⁵⁰ whether all types of IRMAs can progress to neovascularization and whether neovascularization must proceed through precursor IRMA lesions are still debated.^51,52 In the current study, we identified arteriovenous origins of IRMAs consistent with histologic studies. Our PDR eyes showed multiple IRMAs that varied in arteriovenous assignment, with some venular-side IRMAs and others appearing as dilated capillary loops (Fig. 8). Additional investigation into the longitudinal progression of IRMAs that differ in arteriovenous origin may provide further insight toward predicting which IRMAs develop into neovascularization. 

MacTel is a putatively neurodegenerative retinal condition characterized by a constellation of vascular findings including right-angle venules, telangiectatic deep capillaries, and abnormalities of the outer retinal vasculature, which may progress to active neovascularization in some eyes.^53,54 The exact identity of MacTel outer retinal vascular abnormalities, which occupy the normally avascular outer retina, is debated. Some groups consider these lesions to be neovascularization,⁵⁵ whereas others suggest that they are a continuation of the DCP.⁵⁶ In the MacTel eye, we visualized two areas where diving superficial venules converged onto an outer retinal vascular abnormality. The arteriovenous distribution of the outer retinal vascular abnormality resembled that of the DCP, suggesting that these networks are significantly interconnected. Our group has previously shown an intricate and complex relationship between outer retinal telangiectasias and photoreceptor loss with progression of MacTel.^57,58 As MacTel progresses, the photoreceptors degenerate, with retinal thinning in those areas.^56,59 DCP telangiectasias and subsequent retinal deformation in areas of local photoreceptor loss and outer retinal degeneration may contribute to the appearance of outer retinal vascular abnormalities in MacTel. We cannot definitively rule out angiogenic transformation as a contributing factor in the development of these lesions, which deserves further study in longitudinal datasets. 

Several limitations of our study must be considered, including the small sample size, which was largely dictated by the strict image quality requirement and narrow age criteria for the healthy controls. Although we used MaxEntropy thresholding, which has been used in prior OCTA studies to standardize identification of SCP large vessels,^60,61 the limited field of view of OCTA prevented us from categorizing large SCP vessels by branching order, which would have allowed for more objective classification of arterioles and venules. The tendency of OCTA to overestimate vessel diameter due to its limited lateral resolution (less than the width of a capillary⁶²) precluded the application of diameter-based criteria for distinguishing capillaries from larger vessels.^63,64 Although OCTA is capable of detecting flow signal in vessels of a diameter below the optical resolution limit of the device, thus enabling visualization of capillaries, in theory it would not be possible to resolve flow signal from very closely spaced vessels, another limitation of this study. Although our algorithm achieved >90% accuracy in identifying DCP vortices, errors in capillary assignment were still present in images with large projection or motion artifacts. Furthermore, the accuracy of vortex identification does not directly reflect the accuracy of identifying arteriolar connections. The potential influence of projection artifacts must also be considered, as decorrelation tails underneath large vessels may lead to overestimating MCP or DCP vessel density in those regions. To mitigate this consideration, we used projection-removed images for analysis. We would also like to point out that projection artifacts would be expected to disproportionately overestimate the contribution of the larger diameter venular-connected vessels rather than arteriolar. Therefore, we believe it is unlikely that our results are driven by projection artifacts. Another theoretical limitation of the watershed tool would be the high rate of failure of capillary assignment in eyes with low signal or diffuse capillary loss, which would have to be further tested in the future. Incorporation of OCT structural data, high-resolution scanning protocols,⁶⁵ or image quality improvement through multiple image registration and averaging⁶⁶ could potentially improve the performance of this tool, especially in pathologic eyes with non-perfused capillaries and increase overall reliability of DCP measurements. 

In conclusion, we have presented a method for characterizing and quantifying volumetric arteriovenous relationships in the retinal vasculature. We have demonstrated that the healthy macular capillary plexuses are characterized by arteriolar predominance in the MCP and relative slowing of arteriolar compared to venular flow in the deeper vascular layers. We have also shown that this algorithm effectively identifies connectivity to retinal vascular pathology in diseased eyes, which deserves further study. 

Acknowledgments

Supported by a grant from the National Institutes of Health (R01 EY31815 to AAF). Research instrument support was provided by Optovue, Inc. (Fremont, CA, USA). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. 

Disclosure: J.X. Ong, None; G.O. Bou Ghanem, None; P.L. Nesper, None; J. Moonjely, None; A.A. Fawzi, None