Volumetric Measures of Capillary Nonperfusion on Optical Coherence Tomography Angiography Detect Early Ischemia in Diabetes Without Retinopathy

Abstract

Purpose: The purpose of this study was to compare volumetric 3-dimensional (3D) against standard 2-dimensional (2D) measurements of ischemia for distinguishing early stages of diabetic retinopathy (DR) using optical coherence tomography angiography (OCTA).

Methods: This cross-sectional study considered 82 eyes of 82 patients (aged 51.0 ± 11.9 years) including 27 healthy controls, 31 patients with diabetes mellitus (DM) without DR, and 24 patients with mild nonproliferative DR (NPDR). Using OCTA, we obtained 2D scans and 3D volumes of the superficial capillary plexus (SCP), middle capillary plexus (MCP), and deep capillary plexus (DCP). We calculated geometric perfusion deficits (GPDs), which define ischemic regions as those located farther than a specified threshold distance from the nearest blood vessel. For the GPD parameter, we compared the performance of a 20 µm versus 30 µm cutoff.

Results: On 2D scans, eyes with mild NPDR had significantly higher GPDs in all 3 retinal capillary layers, indicating worse ischemia, compared with both healthy controls and patients with DM without DR, using either threshold (20 µm or 30 µm) to define GPD (all P < 0.05). DM without DR showed no significant difference from healthy eyes in 2D images. Interestingly, however, using 3D volumes, DM without DR eyes had significantly greater DCP GPDs than healthy eyes using a GPD threshold of 20 µm (P = 0.012), but not with 30 µm (P = 0.057).

Conclusions: Using a stringent threshold (20 µm), volumetric OCTA imaging detects significant DCP perfusion defects in diabetic eyes even before DR onset, whereas traditional 2D OCTA does not. Volumetric scans may therefore be more sensitive to early ischemia in diabetes.

Diabetic retinopathy (DR) is the leading cause of blindness in adults around the world.¹ Diabetes is associated with loss of vascular supporting cells resulting in ischemia and breakdown of the blood-retinal barrier, which eventually progress to vision-threatening complications such as diabetic macular edema (DME) and proliferative DR. Clinically, the onset of DR is defined by visible sequelae of microvascular damage on fundoscopic examination such as microaneurysms and dot-blot hemorrhages.^2,3 However, the underlying pathologic processes begin long before these features appear, making detection and risk-stratification of those with diabetic eye disease paramount.⁴ Optical coherence tomography angiography (OCTA), which provides noninvasive, depth-resolved visualization of retinal microvasculature, may allow us to characterize features of diabetic eye disease that precede clinical retinopathy. For example, healthy eyes under hyperglycemic conditions and eyes of patients with diabetes without DR have been suggested to exhibit reversed neurovascular coupling responses to dark adaptation.^5,6 Some studies suggest DM without DR eyes show decreased vessel density compared with healthy eyes,^7,8 whereas others have found no difference⁹ or even a slight increase.¹⁰ 

Although OCTA's capacity for depth resolution enables separate assessment of the three retinal capillary layers—superficial capillary plexus (SCP), middle capillary plexus (MCP), and deep capillary plexus (DCP)—most prior OCTA studies have used en face images that flatten the layer of interest into a 2-dimensional (2D) image. However, each retinal capillary plexus varies in thickness by location, forming multi-layered nets in the parafovea that collapse into a single layer in the perifovea.^11,12 As a result, quantifying nonperfusion from 2D images fundamentally relies on measuring intercapillary spaces that are artificially constructed by flattening vessels into a single planar network. En face images do not distinguish between directly connected and overlapping vessels, and do not account for the contribution of retinal thickness to nonperfusion. 

A small number of studies have attempted to use volumetric 3-dimensional (3D) scans to characterize microvascular changes in diabetes, finding that vessel density decreased with increasing severity of DR.^13–15 These studies did not separately consider the MCP and DCP, which may overlook layer-specific changes. Additionally, the use of vessel density in 3D may be a flawed measure of perfusion, as projection artifacts would disproportionately increase vessel volume and underestimate nonperfusion in the axial dimension.¹⁶ One way to address this issue is to define nonperfusion by setting a threshold distance from the vessel center, identified by skeletonizing the vessels. This concept was first described by Zhang et al. who labeled these regions “avascular areas,”¹⁷ and subsequently, Chen et al. who coined the closely related term “geometric perfusion deficits” (GPDs).¹⁸ Defining ischemia by this method not only reduces the variation introduced by artifactual capillary breaks and vessel diameter, but also overcomes projection artifacts when applied to measuring nonperfusion from volumetric images. A recent study by Zang et al. extrapolated this approach to define ischemia using volumetric OCTA and achieved high sensitivity for distinguishing patients with diabetes from healthy controls, a promising result.¹⁹ Their study included a pool of patients with a wide range of DR severity up to proliferative DR, but did not focus on the earliest changes in the transition to DR. 

In this study, we used GPD to assess nonperfusion, and compare the effectiveness of volumetric 3D against standard 2D images for distinguishing early stages of diabetic eye disease. We hypothesized that 3D images would be more sensitive to the earliest evidence of nonperfusion, by accounting for the axial component of nonperfusion. 

Materials and Methods

This prospective study included healthy controls and patients with a diagnosis of diabetes mellitus (DM) who underwent OCTA imaging from January 2021 to June 2023 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. A written informed consent was obtained from all subjects. 

We included subjects who were 25 to 70 years of age. For patients with diabetes, color fundus photographs were obtained using the Ultra-widefield Scanning Laser Ophthalmoscope (Optomap Panoramic 200; Optos PLC, Dunfermline, Scotland, UK) to confirm staging of DR using the International Clinical Diabetic Retinopathy Disease Severity Scale.³ Only eyes with DM without DR and those with mild nonproliferative DR (NPDR) were considered for this study. Fundus photographs were graded by two masked graders (authors N.L.D. and D.C.C) and disagreements in grading were decided by a third senior grader (author H.F.). We excluded eyes with a history of laser treatment or intra-vitreal injections, significant media or lens opacities, previous intra-ocular surgery not including cataract extraction, and other retinal or anterior-segment disease. We also excluded eyes with center-involving DME, defined as central macular thickness greater than 320 µm for men and 305 µm for women.²⁰ 

Demographic and clinical data including age, sex, axial length, type and duration of diabetes, and most recent hemoglobin A1c (HbA1c; within 6 months) were obtained from the electronic medical records. 

OCTA Imaging

OCTA images were obtained using the RTVue-XR Avanti system (Optovue Inc., Fremont, CA, USA) with split-spectrum amplitude-decorrelation angiography (SSADA) algorithm (version 2017.1.0.151).²¹ The A-scan rate was 70,000 scans/s, using a light source centered on 840 nm with a bandwidth of 45 nm, and an approximately 3 mm × 3 mm area centered on the fovea was captured for each scan. We excluded images with quality index (Q-score) < 8 or significant motion or shadow artifacts. If both eyes of the same patient were eligible according to study criteria, we included the eye with the highest Q-score, so that only one eye was enrolled for each study subject. 

The SCP, MCP, and DCP were segmented as previously described,²² with the SCP 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 (OPL). Standard 2D scans were exported with projection artifacts removed using the built-in Optovue software. For volumetric analysis, as it was not possible to export projection-removed 3D volumes from Optovue software, we exported raw 3D volumes and removed projection artifacts using a custom MATLAB program based on the method used by Zhang et al., as previously described.^23,24 Images were inspected for segmentation errors; none in this particular dataset were identified that required manual correction of segmentation. 

Image Analysis

All image analysis was performed using FIJI, an open-source distribution of the program ImageJ.²⁵ Because the axial resolution of the device is approximately 3 µm, whereas the lateral resolution is 10 µm, the images were resized to equalize scaling in the axial and lateral dimensions, without interpolation to avoid introducing new information. The 2D images then were binarized using global Huang thresholding. The 3D images were binarized separately by layer using Otsu thresholding. The layers were separately binarized to limit the influence of signal attenuation with depth that would occur if the full volume were binarized using a single threshold. Both 2D and 3D binarized scans were then despeckled to remove regions of noise that were not connected to surrounding vessels. The vessels were skeletonized to a single pixel width. In order to avoid overestimation of ischemia in areas bordering the SCP large vessels, the SCP large vessels were not skeletonized but instead binarized (using MaxEntropy thresholding), de-noised by applying a circularity cutoff, and reconnected using a method we have previously described.²⁴ The SCP large vessel map was then recombined with the skeletonized capillary map (see Supplementary Fig. S1). 

GPD was calculated using the approach described by Chen et al. that we modified to consider two different thresholds for defining nonperfusion, strict and lenient.¹⁸ The strict threshold defined nonperfusion as farther than 20 µm from the nearest vessel, whereas the lenient threshold defined nonperfusion as farther than 30 µm from the nearest vessel. GPDs were expressed as the percentage of the nonperfused area of the total scan area for 2D images, or as the percentage of the nonperfused volume over retinal slab volume for 3D images (Fig. 1). The foveal avascular zone (FAZ) was manually traced from the MCP slab. The FAZ area or its corresponding cylinder were removed from the total scan area or volume, respectively, before performing any GPD calculations. 

Figure 1.

 

Determining geometric perfusion deficits (GPDs) from standard 2D and volumetric 3D scans. Top row (A–C): Standard 2D volumetric scans of the superficial (SCP), middle (MCP), and deep capillary plexuses (DCP), respectively, were binarized and skeletonized (middle row, D–F) to identify vessels (white) and GPDs were defined as areas either > 20 µm (strict threshold; yellow) or > 30 µm (lenient threshold; red) from the nearest vessel. Volumetric scans were processed in the same way and example B-scans through the SCP, MCP, and DCP, respectively, are shown in (G) through (I).

Statistical Analysis

Statistical analysis was performed using IBM SPSS Statistics version 29 (IBM SPSS Statistics; IBM Corporation, Chicago, IL, USA). Demographic and clinical data were compared with independent 2-sample t-tests or ANOVA tests for continuous data, and Pearson's chi-square tests for categorical data. To compare GPDs between healthy, patients with DM without DR, and mild NPDR groups, we used generalized linear models adjusted for age followed by post hoc Bonferroni correction to account for multiple pairwise comparisons between groups. Receiver operating characteristic (ROC) analysis was also performed for parameters that showed significant differences between healthy controls and patients with DM without DR groups to determine sensitivity and specificity. 

Results

A total of 27 healthy controls and 55 patients with diabetes (31 DM without DR and 24 with mild NPDR) met the inclusion criteria for this study. The three groups were significantly different with respect to age (P = 0.014), the healthy controls being younger compared with the patients with diabetes. The patients with DM without DR group, on average, had a shorter duration of diabetes than the patients with mild NPDR (P = 0.003), but the groups had otherwise similar average HbA1c and relative proportion of type 1 to type 2 diabetes (Table 1). There were no significant differences in axial length (P = 0.800) between the groups. 

Table 1.

 

Demographic and Clinical Characteristics of Study Patients

 

On univariate analysis (Table 2), using a strict 20 µm threshold, GPDs in all layers (SCP, MCP, and DCP) were significantly different between the groups (P ≤ 0.008), whether using standard 2D or volumetric 3D images. Using a lenient 30 µm threshold, all GPDs (P ≤ 0.007), except volumetric SCP GPD, were significantly different (P = 0.072). Post hoc comparisons between groups, adjusting for age, were then performed to elucidate these differences in GPD (Fig. 2). Using 2D images, we found a significant increase in GPD in all layers in patients with mild NPDR compared with both healthy controls and DM without DR eyes (all P < 0.05), but no significant differences between healthy controls and DM without DR eyes (Fig. 3). 

Table 2.

 

Univariate Analysis of Geometric Perfusion Deficits Calculated From 2-Dimensional Scans and 3-Dimensional Volumes in Patients With and Without Diabetes

 

Figure 2.

 

Comparisons of geometric perfusion deficits (GPDs) in the superficial (SCP), middle (MCP), and deep capillary plexuses (DCP) between healthy, diabetic without retinopathy (DM without DR), and mild nonproliferative DR (mild NPDR) groups. Data were adjusted for age. Rows: (Top, A and B) GPDs calculated with a strict threshold. (Bottom, C and D) GPDs calculated with a lenient (30 µm) threshold. Columns: (Left, A and C) Standard 2D scans. (Right, B and D) Volumetric 3D scans. All data were adjusted for age. * P < 0.05; ** P < 0.01.

Figure 3.

 

Standard 2D OCTA scans show worse ischemia in all 3 capillary layers for eyes with diabetic retinopathy (DR) but not in diabetic eyes without DR compared with healthy controls. Rows: Superficial (SCP; top A–C), middle (MCP; middle, D–F), and deep capillary plexuses (DCP; bottom, G–I). Columns: Example eyes from a 40-year-old male healthy control (left), a 40-year-old male patient with diabetes but no DR (middle), and a 51-year-old male patient with mild nonproliferative DR (right). Images have been processed and color-coded to show geometric perfusion deficits (GPDs). White = vessel skeletons, blue = perfused areas within 20 µm of vessel, yellow = areas between 20 to 30 µm from nearest vessel (strict GPD threshold), red = > 30 µm from the nearest vessel (lenient GPD threshold).

In contrast, using a strict 20 µm threshold for defining GPD in volumetric 3D images revealed significantly higher DCP GPD in DM without DR compared with healthy control eyes (see Fig. 2; P = 0.012). This difference was only noted in the DCP; SCP and MCP GPDs were significantly different in mild NPDR eyes relative to healthy and DM without DR eyes. No significant difference between healthy and DM without DR eyes was identified when GPD was defined using the lenient 30 µm threshold (P = 0.057). 

ROC analysis was performed to evaluate the ability of volumetric DCP GPDs to distinguish healthy from DM without DR eyes (Fig. 4). This metric showed good performance with an AUC of 0.743, sensitivity 84%, and specificity 63%. 

Figure 4.

 

Receiver operating characteristic curve of 3D geometric perfusion deficits (GPD) in the deep capillary plexus (DCP) for distinguishing healthy controls from diabetes without retinopathy (DM without DR). DCP GPDs calculated from 3D scans using a strict (20 µm) threshold were significantly increased in DM without DR eyes compared with healthy controls. ROC analysis was performed to assess the ability of this parameter to distinguish the two groups. AUC = area under curve; SN = sensitivity; SP = specificity.

Discussion

In this study, we investigated the relationship between nonperfusion on volumetric 3D as compared with standard 2D OCTA images in early stages of diabetic eye disease. Volumetric images demonstrated significantly higher DCP ischemia in diabetic eyes, even before the onset of clinical retinopathy, which was not captured on conventional 2D images (see Figs. 2, 5). This early change in the DCP was best seen when nonperfusion was defined using a stricter GPD threshold (20 µm from the nearest vessel) rather than the more lenient 30 µm cutoff used in previous studies (see Fig. 2).¹⁸ 

Figure 5.

 

Volumetric scans capture early ischemic areas in eyes of subjects with diabetes without retinopathy that are missed or underestimated in standard 2D images (indicated by green asterisks corresponding to white dotted circles on B-scans). Images have been processed and color-coded to show geometric perfusion deficits (GPDs). White = vessel skeletons, blue = perfused areas within 20 µm of vessel, yellow = areas between 20 to 30 µm from nearest vessel (strict GPD threshold), red = > 30 µm from nearest vessel (lenient GPD threshold). (A, D) The 2D OCTA of deep capillary plexus of 63-year old female patient (A) and a 30-year-old male patient (D), both with diabetes but no retinopathy. (B, C) Examples of volumetric measurement of GPDs on example B-scans corresponding to upper and lower dashed lines, respectively, for patient in A. (E, F) Example B-scans corresponding to upper and lower dashed lines, respectively, for patient in D.

Our study is unique in the focus on early stages of diabetic eye disease, specifically DM without DR and mild NPDR, our analysis of nonperfusion in all three macular capillary layers, and the use of two thresholds for ischemia and volumetric GPDs. Previous volumetric OCTA studies analyzed a wide range of DR as a single group^13,14 or studied patients with type 1 diabetes only.¹⁵ These prior studies measured aggregated retinal nonperfusion without considering individual capillary plexuses, limiting their ability to distinguish changes that may start in specific capillary layers. Recently, Zang et al. showed that volumetric measurement of ischemia achieved high accuracy in distinguishing healthy from diabetic eyes.¹⁹ Differently from our study, these authors considered a pooled sample of eyes with a wide range of DR severity from DM without DR to proliferative diabetic retinopathy (PDR), limiting their ability to resolve changes that might be specific to earlier versus later stages of DR. Although their basing ischemia on diffusion from the center of a vessel is conceptually similar to our GPD calculation, they defined ischemia as 2.5 standard deviations greater than a normative group of 17 healthy eyes that was, on average, 15 years younger than their diabetic groups. This method raises a concern regarding the relative influence of diabetes versus age on their metrics. To address age as a potential confounder in our study, we adjusted for age in our post hoc statistical comparisons of GPDs between diabetes severity groups. 

Our finding of DCP nonperfusion as the earliest change in DM without DR eyes is consistent with previous studies that used standard OCTA approaches to show higher deep capillary ischemia in DM without DR eyes compared with healthy eyes^26,27 and longitudinal studies that show progressive worsening of nonperfusion in DM without DR eyes over time.⁷ Our study advances the field by showing that nonperfusion specifically starts in the DCP, but not the MCP, in early diabetes before the onset of DR. We have also previously shown that deep capillary ischemia distinguished eyes with high risk DR, as well as predicted those that would develop vision-threatening complications of DR.^28,29 The photoreceptors, as the most metabolically active cells in the retina, are spatially closest to the DCP, which contributes to their blood supply especially in dark adaptation.^30,31 At the histologic level, diabetes is associated with mural cell loss before the onset of DR, as well as inflammation and leukostasis resulting in capillary closure.^4,32 Diabetes is also associated with impaired autoregulation of retinal blood flow.^33,34 We have shown that hyperglycemia in healthy subjects as well as patients with diabetes before the onset of DR is associated with a paradoxical constriction of the DCP in dark adaptation, which may further exacerbate ischemia.^5,6 Whether DCP nonperfusion reflects structural capillary closure, or transient nonperfusion due to impaired autoregulation in these eyes with diabetes before the onset of DR, is an important question that needs to be addressed. 

Several factors may contribute to the higher accuracy of volumetric imaging for identifying ischemic changes compared to standard 2D imaging. Whereas 2D images compress all vessels within a given capillary layer into a single plane, 3D images preserve the spatial relationships between vessels and can distinguish vessels that overlap from vessels that directly connect to each other. As a result, volumetric imaging also visualizes perfusion in the axial dimension as well as the lateral, revealing ischemic regions that may otherwise be missed or underestimated on 2D imaging (see Fig. 5). Interestingly, a handful of prior 2D studies showed that early changes in diabetes were most significant in the peri-foveal vascular network.^27,35 Although these findings may truly be related to preferential damage to the perifovea in diabetes, another potential explanation is that the capillaries networks fuse into a single planar layer in the perifovea such that 2D images would closely represent the 3D capillary network in that region.^11,12 In contrast, the presence of axially interlacing capillary layers outside the parafovea would then lead to apparent underestimation of the intercapillary spaces on collapsed, 2D images. 

GPD has several advantages compared with other previously used metrics like vessel density, where perfusion is quantified as percentage of scan volume occupied by vessels, making these measurements susceptible to artifactual variations in vessel width.^13,15 By defining nonperfusion based on the radial diffusion of oxygen from the center of a blood vessel, GPD is resilient to a few notable artifacts. For example, OCTA overestimates capillary width relative to large vessels.³⁶ Additionally, projection artifacts stretch the apparent vessel diameter in the axial direction due to decorrelation tails, which would artifactually assign perfusion to areas within the decorrelation tail.¹⁶ This is an especially important consideration in volumetric imaging, where axial nonperfusion would provide a unique perspective. By defining ischemia as a set distance from the center of a vessel, GPDs limit the contributions of vessel diameter variation and projection artifacts, as well as the influence of artifactual vessel breaks due to noise.^17,18 

We found that a stricter 20 µm GPD threshold may better identify early ischemia or at-risk areas in the early stages of diabetic eye disease. Chen et al., in defining GPD, considered ischemia to be regions more than 30 µm away from the nearest vessel center.¹⁸ Importantly, the 30 µm cutoff is not a physiologic value, but rather an estimate based on the normative foveal intercapillary distance in prior studies of healthy eyes.³⁷ Notably, the Krawitz et al. study that measured this average intercapillary distance used a patient population that was, on average, 52 years old. Capillary density not only decreases with ischemia, but also with age.³⁸ It is possible that a 30 µm threshold may miss ischemia in the eyes of younger patients. Additionally, the intercapillary distance of flow voids directly adjacent the fovea may not be representative of flow voids in the parafovea, where the macular capillaries become multi-layered. We have shown that the 30 µm GPD cutoff had high sensitivity for distinguishing eyes with clinical DR³⁹ and predicting those at risk of developing vision-threatening complications associated with progression to PDR.²⁸ Notably, clinically referable eyes are more ischemic at baseline and would be expected to have larger GPD overall. It is reasonable to speculate that the threshold for ischemia may need to be adjusted in different studies depending on DR severity and the clinical questions being studied. We acknowledge that the chosen value of these distance thresholds is subjective and will be limited by the resolution of the OCTA device—for example, our thresholds were in increments of 10 µm due to the device's lateral resolution of 10 µm per pixel. 

We found no differences in SCP GPD between healthy eyes and DM without DR eyes regardless of whether 2D or 3D images were used. Interestingly, studies by Rosen et al. and Onishi et al. have found DM without DR eyes have higher perifoveal capillary density and SCP flow, respectively, suggesting that capillary perfusion changes in diabetes may involve flow redistribution between the different capillary layers.^10,40 One potential explanation of these differences may be related to the current study's use of GPD as a nonperfusion metric rather than vessel density. Because calculating GPD involves skeletonization of the vessels, this approach may miss perfusion deficits related to variations in vessel caliber. In support of this assumption, imaging modalities such as Doppler and adaptive optics scanning laser ophthalmoscopy (AOSLO) have shown that prior to DR onset, retinal vasculature shows increased capillary flow rates,⁴¹ loss of vascular mural cells,⁴² and arteriolar wall thickening.⁴³ Studies that correlate OCTA perfusion with vascular structure and blood flow quantification could be useful in this respect. 

Limitations to this study include long analysis time, and large computational requirements for volumetric image processing. Because 3D images are composed of stacks of 2D images, processing time, and space required to store the images increase exponentially. For example, image processing and analysis for each 3D volume took 3 to 4 minutes, compared with less than a second for a standard 2D image. Support for 3D image processing is also not built into commercial OCTA software, which limits its application in a clinical setting, but may become feasible with improvements in computer processing power. ROC analysis showed that volumetric DCP nonperfusion, although significantly different between healthy and DM without DR eyes, was only moderately (84%) sensitive in identifying eyes with DM without DR, suggesting that a substantial proportion of DM without DR eyes do not have any detectable ischemia in either 2D or 3D. The cross-sectional design of this study also does not address whether ischemic regions in eyes continue to expand over time, or if they predict clinical consequences like DR progression. 

In conclusion, we found that 3D OCTA imaging detects significant DCP perfusion defects in diabetic eyes even before the onset of retinopathy, findings that are not detectable on traditional 2D OCTA. The 3D volumetric scans of the retina may therefore be more sensitive to early ischemia in diabetes. Further longitudinal studies are warranted to characterize the longitudinal progression and prognostic significance of these early ischemic areas. 

Acknowledgments

AAF was funded in part by NIH grant R01 EY31815. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. 

Commercial Relationships: Research instrument support was provided by Optovue, Inc., Fremont, California, USA. Optovue, Inc. had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. 

Disclosure: J.X. Ong, None; H.J. Lee, None; N.L. Decker, None; D. Castellanos-Canales, None; H. Fukuyama, None; A.A. Fawzi, Genentech/Roche (C), Regenxbio (C), 3Helix (C), Boehringer Ingelheim (C, F)