Deep Capillary Geometric Perfusion Deficits on OCT Angiography Detect Clinically Referable Eyes with Diabetic Retinopathy

Abstract

Purpose:

To evaluate the sensitivity (SN) and specificity (SP) of optical coherence tomography angiography (OCTA) parameters for detecting clinically referable eyes with diabetic retinopathy (DR) in a cohort of patients with diabetes mellitus (DM).

Design:

Retrospective, cross-sectional study.

Subjects:

Patients with DM with various levels of DR.

Methods:

We measured vessel density, vessel length density (VLD), and geometric perfusion deficits (GPD) in the full retina, superficial (SCP) and deep capillary plexus (DCP) on 3 × 3 mm OCTA images. GPD was recently described as retinal tissue located further than 30 μm from blood vessels, excluding the foveal avascular zone (FAZ). We modified the GPD metric by including the FAZ for an additional variable (GPDf). Clinically referable eyes were defined as moderate nonproliferative DR (NPDR) or worse retinopathy and/or diabetic macular edema (DME). One eye from each patient was selected for the analysis based on image quality. We used a binary logistic regression model to adjust for covariates.

Main Outcome Measures:

SN, SP, and area under the receiver operating characteristic curve (AUC).

Results:

Seventy-one of the 150 included eyes from 150 patients (52 DM without DR, 27 mild NPDR, 16 moderate NPDR, 10 severe NPDR, 30 proliferative DR, and 15 eyes with DME) had clinically referable DR. GPDf performed better than GPD in detecting referable DR in the SCP (P = 0.025), but not the DCP or full retina (P > 0.05 for both). DCP GPD had the largest AUC for detecting clinically referable eyes (AUC = 0.965, SN = 97.2%, SP = 84.8%), which was significantly larger than the AUC for vessel density of any layer (P < 0.05 for all), but not DCP VLD (P = 0.166). The cut-off value of 2.5% for DCP GPD resulted in a highly sensitive test for detecting clinically referable eyes without adjusting for covariates (AUC = 0.955, SN = 97.2%, SP = 79.7%).

Conclusions:

Vascular parameters in OCTA, especially in the DCP, have the potential to identify eyes that warrant further evaluation. GPD may better distinguish these clinically referable DR eyes than standard vessel density parameters.

Précis:

Geometric perfusion deficits in the deep capillary plexus on optical coherence tomography angiography had a high sensitivity and performed better than vessel density in identifying diabetic retinopathy eyes that warrant further evaluation.

Diabetic retinopathy (DR) is a leading cause of visual impairment in adults.¹ DR is a microangiopathy characterized by the degeneration of pericytes and endothelial cells of the retinal capillaries, leading to microaneurysms, vascular occlusion, leakage and diabetic macular edema (DME).^2,3 Progression of ischemia ultimately leads to aberrant neovascularization of the retina and anterior segment, which are causes for significant visual impairment.⁴ The use of retinal imaging has allowed clinicians to detect early pathological changes in the course of the disease and implement treatments, which can slow DR progression and reduce vision loss.^5,6 Based on vascular parameters in color fundus photography, the International Clinical Diabetic Retinopathy (ICDR) established a systematic formula that categorizes DR into different stages, to facilitate recommendations for follow up based on epidemiologic studies.⁷ These stages include diabetes mellitus without DR (DM without DR), mild, moderate, and severe nonproliferative (NPDR), and proliferative (PDR) stages. Grading color fundus photographs in the clinic is a reliable method for accurate classification of DR severity and can now be done using FDA-approved artificial intelligence software.⁸

Optical coherence tomography angiography (OCTA) is a relatively new imaging technique for visualizing and quantitatively evaluating the retinal vasculature in vivo. Most academic centers now have one or more OCTA devices and the high-resolution 3D capability of this technology has led to a greater understanding of the microvasculature of the retina in health and disease.⁹ Yet, the clinical usefulness of these devices for managing patients with DR remains unclear. An important test of the utility of this device is its ability to provide quantitative data that correlate with the severity of DR and distinguish eyes that require immediate treatment or evaluation by a specialist. In DR, eyes with moderate NPDR or worse, along with eyes that have clinically significant DME, are considered at risk and referral to an ophthalmologist is recommended.¹⁰ The Early Treatment Diabetic Retinopathy Study (ETDRS) reported an annual rate of moderate vision loss (15 letters on the ETDRS chart) of 8% in patients with untreated DME.¹¹ Epidemiologic studies also showed that approximately 50% of eyes with severe NPDR will develop some degree of PDR within 1 year and 44% of eyes with high-risk PDR will progress to severe vision loss (worse than 5/200) at 4 years without intervention.^12,13 A more recent study reflecting improved systemic treatment for DM reported that the progression to PDR at 5 years from a baseline of moderate NPDR was 27% and from a baseline of severe NPDR was 46%.¹⁴ Therefore, the International Council of Ophthalmology recommends referral or next visit within 3–6 months for moderate NPDR, <3 months for severe NPDR, <1 month for PDR and 1 month for center-involved DME in high-resource settings, in order to reduce the burden of preventable vision loss in these patients.¹⁰

A number of studies have recently evaluated the correlation of OCTA with severity of DR.^15,16 These studies generally use the ICDR classification to stage the DR and then assess the ability of the different OCTA parameters to classify these eyes into their respective stage. Geometric perfusion deficits (GPD) is an OCTA-based measurement, recently described by Chen et al.¹⁷ that takes into account established oxygen diffusion gradients as well as the predicted intercapillary spacing in the healthy macula based on imaging studies.^18–21 GPD has the potential to highlight retinal zones that are ischemic. Since oxygen is consumed by retinal tissue as it diffuses transversely in the retina, tissues located more than 30 microns from the nearest blood vessel are more likely to have an oxygen tension below the threshold of ischemia.¹⁷ Chen and colleagues found that GPD had significantly better repeatability than vessel density in most comparisons. They also found that GPD was able to distinguish between DR groups with better statistical performance than vessel density, despite their very limited sample size of only 12 eyes with DR.¹⁷ The authors did not include the FAZ as part of the GPD measurement in their study since oxygen diffusion from the choroid supplies this region under normal conditions. Yet, the dropout of capillaries around the FAZ in DR leads to FAZ enlargement and diabetic macular ischemia (DMI), which we believe should be taken into account.^22,23 Based on the preliminary results from the Chen et al. study¹⁷ and the previous studies suggesting the importance of the FAZ, we hypothesized that GPD including the FAZ (GPDf) could be more effective at distinguishing clinically referable eyes than vessel density parameters or FAZ area alone. We discuss the benefits of OCTA used in adjunct with conventional imaging methods as well as the use of the GPD parameter as a visual tool for evaluating patients with DMI.

Methods

This retrospective, cross-sectional study was approved by the Institutional Review Board of Northwestern University, which follows the tenets of the Common Rule of the National Institutes of Health. This study was performed in accordance with HIPAA regulations and adhered to the tenets of the Declaration of Helsinki. Patients were originally recruited for OCTA imaging between July of 2015 and December of 2020 in the Department of Ophthalmology of the Feinberg School of Medicine. We obtained informed consent from each participant before imaging.

Patient Criteria

Patients were eligible if they have a diagnosis of DM based on diagnostic cutoffs for fasting blood glucose and glycated hemoglobin or hemoglobin A1c (HbA1c).²⁴ We only included eyes with an OCTA scan quality (Q-Score) of 7 or higher (out of 10) and minimal motion or shadowing artifacts. Exclusion criteria were eyes with previous ocular surgery or any other retinal disease or suspected disease, such as suspected glaucoma. We also excluded eyes with significant cataract to avoid artifacts that compromise OCTA image quality. Classification of the level of DR, including DM without DR, mild/moderate/severe NPDR and PDR was obtained from chart review only (n = 48) or from chart review and color fundus photographs (n = 102) using the ICDR severity scale.⁷ Any discrepancy between the color photograph grade assigned by first grader (PLN) and the grade assigned in the medical record was resolved through masked grading by a second grader (JXO) as the tie-breaker. Clinically referable eyes were defined as a DR grade of moderate NPDR or higher, or the presence of DME, which was defined as the presence of center-involving cystic spaces or hard exudates as well as a central macular thickness (CMT) of greater than 300 μm.²⁵

OCTA Imaging

We used the RTVue-XR Avanti system (Optovue, Inc., Fremont, CA, USA) with split-spectrum amplitude-decorrelation angiography (SSADA) software version (2017.1.0.151) to obtain 3 × 3 mm (304 pixels × 304 pixels) OCTA images centered on the fovea.²⁶ The specifics of the device have been described previously.²⁷ This device is a modified OCT system that captures the movement of red blood cells in the retina by acquiring two subsequent images of the same location. In a healthy stationary eye, movement signals can be attributed to blood flow, which is transformed by the SSADA algorithm into a 3D reconstruction of the vasculature.²⁶ The final 3 × 3 mm scan centered on the fovea has a depth of 2 mm and contains a total of 304 B-scans, each with 304 A-scans.

Image Analysis

We used the automated AngioVue Analytics software (2017.1.0.151) for segmenting three vascular layers, the full retinal vascular network, the superficial capillary plexus (SCP) and the deep capillary plexus (DCP). The full retina slab was defined as the tissue from the inner limiting membrane (ILM) to 10 μm below the outer plexiform layer (OPL). The SCP was defined as the tissue from the ILM to 10 μm above the inner plexiform layer (IPL) and the DCP defined as the tissue from 10 μm above the IPL to 10 μm below the OPL. We obtained the automated parafoveal vessel density from the AngioVue software for each layer to compare these sensitivities (SN) and specificities (SP) to those from the calculated GPD. The parafovea was defined as an annulus centered on the fovea with inner and outer ring diameters of 1 and 3 mm, respectively. Vessel density was defined as the area of vessel pixels divided by the total area of interest and was reported as a percentage. We calculated vessel length density (VLD) after binarization using Huang2 thresholding and skeletonization of the vessels.²⁸ VLD (mm⁻¹) was defined as the total length of skeletonized vessels divided by the area as previously reported.²⁹ CMT and the presence of cysts or hard exudates were also recorded. Images were then exported into FIJI software.³⁰ In FIJI, the FAZ was manually delineated using the full retinal vascular network slab and measured as an area (mm²).

GPD is defined as the percent area of the retina that is located greater than 30 μm form the nearest blood vessel, based on predicted intercapillary distances in healthy eyes.¹⁷ A semi-automated FIJI macro was used to measure GPD. We binarized each of the three vascular layers using the automated global thresholding method Huang2 in order to convert the grayscale image into a binary white and black image that can be used for quantification.²⁸ Then, the capillaries were skeletonized since the smaller capillaries near the center of the macula are below the transverse resolution of the device. We then binarized the large vessels using the Max Entropy plugin.³¹ The final image for GPD quantification was an overlay of Max Entropy binarized large vessels and skeletonized capillaries (Figure 1). Speckle and noise were manually eliminated before calculation of GPD. We calculated the GPD percent for each vascular layer excluding the FAZ (GPD) and including the FAZ (GPDf).

Figure 1. Image processing and measurement of geometric perfusion deficits (GPD).

(A) Original gray-scale OCTA image of the superficial capillary plexus (SCP). (B) SCP after Huang2 global thresholding. (C) SCP after Huang2 image was skeletonized. (D) SCP after Max Entropy thresholding to delineate the large SCP vessels. (E) Overlay of skeletonized capillaries from (C) and binarized large vessels from (D). This image was used to calculate the GPD including the foveal avascular zone (GPDf), which is tissue greater than 30 microns from the nearest blood vessel and represented as red pixels in (F).

Statistics

We used SPSS version 27 (IBM SPSS Statistics; IBM Corporation, Chicago, IL, USA) to perform statistical tests. Shapiro-Wilk tests were significant for GPD and GPDf, suggesting these parameters did not follow a normal distribution. Shapiro-Wilk tests were not significant for vessel density or VLD, suggesting normal distributions. We compared the mean and standard deviation of demographic and outcome variables between the referable and non-referable groups using the nonparametric independent-samples Mann-Whitney U tests for non-normal data and parametric independent samples t-tests for normal data. Two-way, random, absolute agreement intraclass correlation coefficients (ICC) were calculated for GPD as well as the FAZ area measurement. We then performed a binary logistic regression model for each of the outcome variables while adjusting for covariates that were significantly different between referable and non-referable eyes, which included sex, DM type, refractive error, HbA1c, and Q-score. We replaced the missing HbA1c value in 6 subjects and the missing refractive error in 27 subjects with the mean value for their respective DR severity group, which has been shown to improve the power of the model without introducing bias.³² The predicted probability values from the regression models were used to generate receiver operating characteristic (ROC) curves. The sensitivity, specificity, and area under the curve (AUC) for detecting clinically referable eyes was calculated for GPD, GPDf, vessel density, VLD, and FAZ area. The differences in AUC were compared using the “paired-sample area difference under the ROC curve” in SPSS, which based on the nonparametric method proposed by DeLong and colleagues.³³

We generated a separate binary logistic regression model using only treatment naïve eyes, which included sex, DM type, HbA1c, and Q-score as covariates, since refractive error was no longer significant in the treatment naïve group. We calculated ROC curves for referable treatment naïve eyes. We also sought to provide readers with a GPD value that could be used as a cut-off for determining whether a patient should be referred to a specialist. Since binary logistic regression models result in a probability value instead of a GPD percentage, which is subsequently used to generate the ROC curve, we generated ROC curves from the raw GPD and vessel density data to provide cut-off values. We then used a Partial Correlation model correcting for significant covariates in the entire cohort and in the treatment-naïve eyes to evaluate the correlation of OCTA parameters with DR severity. We considered a P value of less than 0.05 statistically significant.

Results

Seventy-one of the 150 included eyes had referable levels of DR. Overall, the study eyes included 52 eyes with DM without DR, 27 mild NPDR, 16 moderate NPDR, 10 severe NPDR, 30 PDR, and 15 eyes with DME and various levels of DR. In the DME group, 1 had mild NPDR, 6 had moderate NPDR, 3 had severe NPDR and 5 had PDR. The demographic data for each group are shown in Table 1. Parameters that were significantly different between the referable and non-referable groups in Table 1 were included as covariates in the binary logistic regression model. OCTA parameters and ICC, which showed excellent agreement, are reported in Table 2. An example showing the difference in GPDf (FAZ included) between mild and moderate NPDR is shown in Figure 2.

Table 1.

Diabetic retinopathy severity and demographic characteristics in referable and non-referable eyes

	DM without DR	Mild NPDR	Moderate NPDR	Severe NPDR	PDR	DME	Non-Referable	Referable	P-value
Number of Subjects	52	27	16	10	30	15	79	71
Age (mean ± SD)	49.1 ± 17.0	48.7 ± 10.9	52.6 ± 11.8	52.1 ± 10.9	46.3 ± 10.3	56.4 ± 12.8	49.0 ± 15.1	50.7 ± 11.7	0.326
Sex, n Female (%)	32 (62%)	16 (59%)	7 (44%)	5 (50%)	11 (37%)	3 (20%)	48 (61%)	26 (37%)	0.003* a
Refractive Error (D mean ± SD)	−2.09 ± 2.36	−1.33 ± 2.63	−1.01 ± 2.33	−0.54 ± 2.84	−0.84 ± 2.46	−0.71 ± 2.19	−1.81 ± 2.47	−0.81 ± 2.35	0.028*
Missing, n (%)	11 (21%)	3 (11%)	2 (13%)	3 (30%)	7 (23%)	1 (7%)	14 (18%)	13 (18%)
DM Type, n Type 1 (%)	22 (42%)	17 (63%)	6 (38%)	0 (0%)	11 (37%)	0 (0%)	39 (49%)	17 (24%)	0.001* a
DM Duration (mean years ± SD)	12.0 ± 9.7	23.5 ± 12.9	22.3 ± 9.5	12.2 ± 4.4	21.2 ± 10.5	13.7 ± 9.1	15.9 ± 12.1	18.5 ± 10.0	0.069
HbA1c (mean ± SD)	7.2 ± 1.3	7.6 ± 1.5	8.3 ± 1.6	8.7 ± 2.1	8.6 ± 1.8	8.2 ± 1.6	7.4 ± 1.4	8.5 ± 1.7	<0.001*
Treatment, n (%)	0 (0%)	0 (0%)	2 (13%)	1 (10%)	19 (63%)	10 (67%)	0 (0%)	32 (45%)	<0.001* a
Q-Score (mean ± SD)	8.3 ± 0.7	8.4 ± 0.6	8.0 ± 0.8	8.3 ± 0.7	7.9 ± 0.7	8.1 ± 0.5	8.3 ± 0.7	8.0 ± 0.7	0.008*

Differences in demographic characteristics between the referable and non-referable groups were tested using nonparametric independent-samples Mann-Whitney U tests unless noted with (^a), which signifies independent-samples t-tests for demographic data that was normally distributed. P-values were obtained by comparing non-referable to referable groups. Non-referable included DM without DR and mild NPDR, while referable included moderate and severe NPDR, PDR and DME. DME = diabetic macular edema, D = diopters, DM = diabetes mellitus, DR = diabetic retinopathy, HbA1c = glycated hemoglobin, NPDR = nonproliferative diabetic retinopathy, PDR = proliferative diabetic retinopathy, SD = standard deviation.

^*Statistical significance (P < 0.05).

^aIndependent-samples t-tests for normally distributed data.

Table 2.

Optical coherence tomography angiography vascular parameters by diabetic retinopathy severity

	DM without DR	Mild NPDR	Moderate NPDR	Severe NPDR	PDR	DME	Non-referable	Referable	P-value	ICC
Full GPDf (% mean±SD)	3.9 ± 1.1	5.3 ± 1.5	8.1 ± 4.8	7.7 ± 2.6	10.8 ± 4.5	8.5 ± 1.9	4.4 ± 1.4	9.3 ± 4.1	<0.001*	0.998 (0.988–0.999)
Full GPD (%)	1.6 ± 0.6	2.8 ± 0.9	5.1 ± 3.9	4.4 ± 1.6	6.6 ± 3.8	5.5 ± 1.7	2.0 ± 0.9	5.7 ± 3.3	<0.001*
Full Vessel Density (%)	58.1 ± 3.7	56.3 ± 2.4	53.1 ± 6.3	52.3 ± 4.4	49.9 ± 4.5	50.4 ± 3.0	57.5 ± 3.4	51.0 ± 4.8	<0.001* a
Full Vessel Length Density (mm−1)	23.6 ± 1.3	22.8 ± 1.5	21.1 ± 2.6	21.4 ± 1.2	20.1 ± 2.1	20.5 ± 1.4	23.3 ± 1.4	20.6 ± 2.0	<0.001* a
SCP GPDf (%)	8.5 ± 2.9	14.0 ± 4.6	17.1 ± 9.2	17.0 ± 4.3	21.6 ± 9.1	16.4 ± 4.3	10.4 ± 4.4	18.8 ± 8.0	<0.001*	0.995 (0.992–0.996)
SCP GPD (%)	6.3 ± 3.0	11.7 ± 4.7	14.5 ± 8.8	14.2 ± 3.8	18.1 ± 9.3	13.8 ± 4.3	8.2 ± 4.5	15.9 ± 7.9	<0.001*
SCP Vessel Density (%)	48.2 ± 3.7	42.8 ± 3.9	41.0 ± 7.3	40.3 ± 4.2	38.1 ± 5.7	40.1 ± 4.1	46.4 ± 4.5	39.5 ± 5.6	<0.001* a
SCP Vessel Length Density (mm−1)	18.6 ± 2.0	16.0 ± 1.9	15.4 ± 3.0	15.1 ± 1.2	14.1 ± 2.6	15.4 ± 1.7	17.7 ± 2.4	14.8 ± 2.4	<0.001* a
DCP GPDf (%)	3.7 ± 1.1	5.3 ± 1.6	8.3 ± 5.1	8.6 ± 3.0	11.6 ± 4.6	9.6 ± 4.3	4.3 ± 1.5	10.0 ± 4.6	<0.001*	0.993 (0.979–0.997)
DCP GPD (%)	1.4 ± 0.6	2.8 ± 1.1	5.2 ± 4.3	5.4 ± 2.2	7.4 ± 3.9	6.7 ± 4.2	1.9 ± 1.0	6.5 ± 3.9	<0.001*
DCP Vessel Density (%)	51.2 ± 3.0	48.6 ± 3.3	45.0 ± 4.9	44.0 ± 3.9	42.0 ± 4.0	40.9 ± 3.2	50.3 ± 3.3	42.7 ± 4.3	<0.001* a
DCP Vessel Length Density (mm−1)	23.8 ± 1.7	21.8 ± 1.7	19.7 ± 2.7	18.8 ± 2.4	17.8 ± 2.1	18.1 ± 2.0	23.2 ± 1.9	18.4 ± 2.3	<0.001* a
FAZ (area, mm2)	0.30 ± 0.11	0.33 ± 0.13	0.41 ± 0.18	0.42 ± 0.15	0.56 ± 0.27	0.41 ± 0.12	0.31 ± 0.11	0.47 ± 0.22	<0.001*	0.994 (0.977–0.997)

Differences in vascular parameters between the referable and non-referable groups were tested using nonparametric independent-samples Mann-Whitney U tests unless noted with (^a), which signifies independent-samples t-tests for demographic data that was normally distributed. P-values were obtained by comparing non-referable to referable groups. Non-referable included DM without DR and mild NPDR, while referable included moderate and severe NPDR, PDR and DME. ICC values reported are two-way random absolute agreements and all were significant at P < 0.001. DME = diabetic macular edema, DCP = deep capillary plexus, DM = diabetes mellitus, DR = diabetic retinopathy, FAZ = foveal avascular zone, GPD = geometric perfusion deficits with FAZ excluded, GPDf = GPD with FAZ included, ICC = intraclass correlation coefficient, NPDR = nonproliferative diabetic retinopathy, PDR = proliferative diabetic retinopathy, SCP = superficial capillary plexus, SD = standard deviation.

^*Statistical significance (P < 0.05).

^aIndependent-samples t-tests for normally distributed data.

Figure 2. Geometric perfusion deficits with foveal avascular zone included (GPDf) on OCTA shows increased ischemic area in more severe diabetic retinopathy (DR).

Top Row is mild nonproliferative (NPDR). Bottom Row is moderate NPDR. Red pixels represent the area classified as GPDf. GPDf was calculated as any area located more than 30 microns from the nearest vessel after we binarized large vessels as well as binarized and skeletonized capillaries.

The ROC curves obtained from the binary logistic regression model for distinguishing clinically referable eyes using GPD and GPDf as well as parafoveal vessel density and VLD are shown in Figure 3. DCP GPD had the largest AUC for detecting clinically referable eyes (AUC = 0.965 [0.936–0.994], SN = 97.2%, SP = 84.8%; Figure 3 and 4), followed by DCP GPDf (AUC = 0.958 [0.929–0.987], SN = 94.4%, SP = 81.0%) and full retina GPD (AUC = 0.958 [0.928–0.987], SN = 94.4%, SP = 78.5%).

Figure 3. Geometric perfusion deficits (GPD) performed better than vessel density and vessel length density (VLD) in detecting clinically referable eyes.

Receiver operating characteristic (ROC) curves with area under the curve (AUC), sensitivity (SN), and specificity (SP) for perfusion deficits with foveal avascular zone (FAZ) included (GPDf) (1^st Row) and excluded (GPD) (2^nd Row) as well as the parafoveal vessel density (3^rd Row) and full scan VLD (4^th Row). Layers of the retina are shown by column with full retinal thickness segmentation (Left Column), superficial capillary plexus (Middle Column) and deep capillary plexus (Right Column). Deep capillary plexus GPD with FAZ excluded had the greatest area under the curve (Middle Row, Right Column; AUC = 0.965, SN = 97.2%, SP = 84.8%).

Figure 4. Deep capillary plexus (DCP) geometric perfusion deficits (GPD) performed best in distinguishing clinically referable eyes with diabetic retinopathy (DR).

Binarized and skeletonized capillaries of the DCP shown with red pixels representing GPD, which does not include the foveal avascular zone. GPD red pixel area represents tissue area that is likely to be ischemic, which increases with more severe DR. DM = diabetes mellitus, NPDR = nonproliferative diabetic retinopathy, PDR = proliferative diabetic retinopathy, DME = diabetic macular edema.

In each vascular layer (full retina, SCP and DCP), both GPD and GPDf had a larger AUC than the vessel density for that layer and this was significant in the full retinal thickness layer (P < 0.05 for both) and for DCP GPD (P = 0.017), but not the SCP (P > 0.05 for both) and DCP GPDf (P = 0.072). GPD and GPDf also had a larger AUC than the VLD for the respective layer and this was significant in the SCP (P < 0.05 for both) and full retina (P < 0.05 for both), but not DCP (P > 0.05 for both). GPDf had a significantly larger AUC than the GPD in the SCP (P = 0.025), but not the DCP or full retina (P > 0.05 for both). DCP GPD had a significantly larger AUC than any AUC calculated from vessel density (P = 0.006 for full retina, P = 0.001 for SCP, P = 0.017 for DCP). DCP GPD also had a significantly larger AUC than the AUC calculated from VLD for the full retina and SCP (P < 0.001 for both), but this difference was not significant for DCP VLD (P = 0.166). The FAZ area had the lowest AUC (AUC = 0.876 [0.822–0.931], SN = 85.9%, SP = 73.4%).

We wanted to explore whether prior treatment affected the accuracy of the algorithm, so we analyzed these groups separately. Considering the 118 treatment naïve eyes (52 DM without DR, 27 mild NPDR, 14 moderate NPDR, 9 severe NPDR, 11 PDR, and 5 eyes with DME and various levels of DR), 39 had referable disease. In the DME group, 3 had moderate NPDR and 2 had severe NPDR. Overall, and similar to the entire cohort, DCP GPD had the greatest AUC (AUC = 0.956 [0.910–1.000], SN = 97.4%, SP = 77.2%; Figure 5), which performed better than vessel density of the DCP (AUC = 0.921 [0.862–0.979], SN = 92.3%, SP = 72.2%; Figure 5) as well as VLD of the DCP (AUC = 0.946 [0.895–0.997], SN = 97.4%, SP = 81.0%), although these were not statistically significant (P = 0.101 for vessel density and P = 0.278 for VLD). In this group, OCTA parameters were able to distinguish clinically referable eyes with a high sensitivity despite the lower proportion of referable eyes (39 of 118). For the entire cohort as well as the treatment naive cohort, DCP performed the best for distinguishing clinically referable eyes (Figure 3 and 5).

Figure 5. Geometric perfusion deficits (GPD) performed better than vessel density and vessel length density (VLD) in detecting clinically referable eyes in eyes without prior diabetic retinopathy treatment.

Receiver operating characteristic (ROC) curves with area under the curve (AUC), sensitivity (SN), and specificity (SP) for perfusion deficits with foveal avascular zone (FAZ) included (GPDf) (1^st Row) and excluded (GPD) (2^nd Row) as well as the parafoveal vessel density (3^rd Row) and full scan VLD (4^th Row). Layers of the retina are shown by column with full retinal thickness segmentation (Left Column), superficial capillary plexus (Middle Column) and deep capillary plexus (Right Column). Deep capillary plexus GPD with FAZ excluded had the greatest area under the curve (Middle Row, Right Column; AUC = 0.956, SN = 97.4%, SP = 77.2%).

The correlation of OCTA parameters with DR severity was assessed using a Partial Correlation while correcting for covariates that were significantly associated with DR severity, including sex, refractive error, DM type, DM duration, HbA1c, previous treatment, and Q-score for the entire cohort and refractive error, DM duration, and HbA1c for the treatment-naïve group. The results are reported in Table 3. We found that DCP GPDf had the strongest correlation with DR severity in the entire cohort ( r(141) = 0.592 [0.481–0.693], P < 0.05), while DCP VLD had the strongest correlation with DR severity in the treatment-naïve group ( r(108) = 0.683 [0.579–776], P < 0.05).

Table 3.

Optical coherence tomography angiography vascular parameters correlate with diabetic retinopathy severity

	Entire Cohort (n=150)		Treatment-Naive Cohort (n=118)
	Correlation r (95% CI)	P-Value	Correlation r (95% CI)	P-Value
Full GPDf	0.576 (0.450–0.691)	<0.001*	0.652 (0.535–0.765)	<0.001*
Full GPD	0.539 (0.436–0.653)	<0.001*	0.606 (0.510–0.736)	<0.001*
Full Vessel Density	0.460 (0.298–0.624)	<0.001*	0.499 (0.297–0.686)	<0.001*
Full Vessel Length Density	0.455 (0.315–0.569)	<0.001*	0.520 (0.379–0.638)	<0.001*
SCP GPDf	0.545 (0.409–0.659)	<0.001*	0.602 (0.459–0.731)	<0.001*
SCP GPD	0.502 (0.367–0.618)	<0.001*	0.558 (0.414–0.686)	<0.001*
SCP Vessel Density	0.499 (0.363–0.618)	<0.001*	0.539 (0.391–0.667)	<0.001*
SCP Vessel Length Density	0.440 (0.283–0.561)	<0.001*	0.491 (0.350–0.622)	<0.001*
DCP GPDf	0.592 (0.481–0.693)	<0.001*	0.668 (0.564–0.776)	<0.001*
DCP GPD	0.552 (0.451–0.656)	<0.001*	0.635 (0.543–0.776)	<0.001*
DCP Vessel Density	0.546 (0.435–0.662)	<0.001*	0.653 (0.530–0.758)	<0.001*
DCP Vessel Length Density	0.590 (0.484–0.688)	<0.001*	0.683 (0.579–0.776)	<0.001*
FAZ Area	0.422 (0.270–0.558)	<0.001*	0.498 (0.329–0.634)	<0.001*

Partial Correlations were used to evaluate the correlation between vascular parameters and diabetic retinopathy severity while correcting for covariates. Eyes with diabetic macular edema were included based on level of diabetic retinopathy. CI = confidence interval, DCP = deep capillary plexus, FAZ = foveal avascular zone, GPD = geometric perfusion deficits with FAZ excluded, GPDf = GPD with FAZ included, SCP = superficial capillary plexus

^*Statistical significance (P < 0.05).

The ROC curves generated from the entire cohort and treatment-naïve group using raw GPD and vessel density data (without adjusting for covariates) were explored to provide a cut-off GPD value. The average AUC using the raw data was 0.041 ± 0.026 lower than the covariate-adjusted values. DCP GPD had the largest AUC. We found that using a cut-off value of ≥ 2.5% for DCP GPD (FAZ excluded) provided a highly sensitive test for detecting eyes from the entire cohort that should be referred, (AUC = 0.955 [0.926–0.985], SN = 97.2%, SP = 79.7%). The sensitivity and specificity of this cut-off value were similar for treatment naïve eyes (AUC = 0.944 [0.905–0.982], SN = 97.4%, SP = 79.7%). Using a cut-off value of ≤ 48.4% for DCP vessel density produced the following results (AUC = 0.913 [0.867–0.958], SN = 91.5%, SP = 70.9% for the entire cohort) and (AUC = 0.900 [0.842–0.958], SN = 87.2%, SP = 72.2% for treatment naïve eyes). Using a cut-off value of ≤ 21.71 mm⁻¹ for DCP VLD produced the following results (AUC = 0.940 [0.905–0.975], SN = 91.5%, SP = 74.7% for the entire cohort) and (AUC = 0.925 [0.877–0.973], SN = 89.7%, SP = 74.7% for treatment naïve eyes).

Discussion

In the current study, we assessed the ability of OCTA non-perfusion metrics, namely vessel density, VLD, FAZ and the GPD parameter, to distinguish eyes that warrant close follow-up with an ophthalmologist (moderate DR or worse and/or DME). We found that DCP GPD can be a reliable metric to identify eyes that need further evaluation. We used a binary logistic regression model to adjust for significant demographic covariates and found GPD (with and without including the FAZ) had a greater AUC than vessel density and VLD at every capillary layer studied in the entire cohort as well as the treatment naïve subgroup. Removing the FAZ from the GPD resulted in the highest overall AUC (DCP layer), which was significantly higher than the AUC for vessel density of any layer (full retina, SCP, DCP) in the entire cohort (P < 0.05 for all), except for DCP VLD (P = 0.166). These differences were not significant when evaluating the treatment naïve subgroup alone. DCP GPD had the strongest correlation with DR severity in the entire cohort and DCP VLD had the strongest correlation in the treatment-naïve group. We found that using a cut-off value of ≥ 2.5% for GPD DCP provided a highly sensitive test for detecting clinically referable eyes (SN = 97.2%, SP = 79.7% for the entire cohort; SN = 97.4%, SP = 79.7% for treatment naïve eyes).

Since the inception of OCTA, a myriad of parameters have been tested for their ability to stratify levels of DR with varied success.^15,25,34 FAZ and vessel density were among the first OCTA parameters tested.³⁵ The enlargement of the FAZ due to perifoveal microvascular occlusion has long been a critical prognostic factor in eyes with DR.³⁶ FAZ enlargement defines DMI, which, in turn, can lead to vision loss as well as DME and is a surrogate marker for peripheral ischemia and ultimately neovascularization.^22,37 Studies that explored whether OCTA or fluorescein angiography is superior for delineating the FAZ in DR and DME have differed in their conclusion, but have generally shown moderate agreement between the two methods.^38,39 Early OCTA studies showed significantly enlarged FAZ with increasing DR severity, and even in patients with DM without DR.^35,40 Studies showed statistically significant differences in FAZ between DR severity groups,³⁴ but the sensitivity and specificity were not acceptable for distinguishing DR status.²⁵ Similarly, we found a relatively low sensitivity and specificity for FAZ area in the current study; FAZ area generally performed the worst (AUC = 0.876 [0.822–0.931], SN = 85.9%, SP = 73.4%). This was not unexpected and can be explained by the high variability in FAZ size even among healthy individuals.⁴¹ Yet, including the FAZ in the GPD parameter (GPDf) increased the AUC significantly compared to the standard GPD (without FAZ) for the SCP (P = 0.025), but not in the full retina or DCP layers, which both had a larger AUC without the FAZ, although these differences were not significant. We also found that including the FAZ in the GPD measurement resulted in a stronger correlation with DR severity in all layers (Table 3). These findings suggest that the correlation between FAZ enlargement and DR severity on OCTA may be best captured in the SCP and that including the FAZ is beneficial when calculating GPD for the SCP.

Vessel density, VLD and GPD were found to have the greatest area under the curve in the DCP (Figures 3–5), consistent with previous studies showing a preferential disruption of the deep plexus in DR.^34,42 GPD and GPDf performed better than vessel density for the respective retinal layer, while DCP GPD had the largest AUC overall. As described by Chen et al,¹⁷ GPD circumvents the issue of small capillary discontinuity that is common in OCTA. For instance, a region with a small discontinuous capillary may be categorized as having lower vessel density and higher intercapillary area. Yet, this may not reflect tissue ischemia unless the discontinuity is larger than 30 microns. GPD is inherently more likely to reflect hypoxic retinal tissue, which we postulated would correlate with DR severity.⁴³ This may explain why GPD performed better than vessel density in our study. It is important to note that DCP VLD performed well in distinguishing eyes with referable DR (Figures 3 and 5) and had the strongest correlation of any parameter with DR severity in the treatment-naïve group (Table 3). Since the DCP has no large vessels and contains only capillaries, skeletonization of the vessels in this layer followed by a density calculation provides an accurate assessment of the ischemic status that is more difficult to evaluate in superficial layers due to problems with skeletonizing large vessels. While VLD is a less intensive calculation, DCP GPD performed slightly better than VLD in the full cohort correlation with DR severity and in distinguishing referable DR in both the full cohort and treatment-naïve eyes, yet these differences were not significant. We suggest that automated software that shows an overlay of GPD may provide a useful visual tool for understanding the extent and location of ischemic changes in the setting of DMI.

While the GPD parameter offered statistically significant improvement in classifying eyes with referable DR compared to vessel density, the comparison lost statistical significance when we focused on treatment naïve eyes. In a recent study, our group showed that both DCP vessel density and DCP GPD had predictive value for complications (i.e., vitreous hemorrhage, DME, treatment initiation) occurring within 1 year, but there were no significant differences between vessel density and GPD (J. X. Ong, et al., 2022, In review). Together, these studies suggest that since the automated DCP vessel density measurement from AngioVue software is readily available, it may be a more practical parameter that can capture clinically referable eyes and predict clinical complications. It is worth noting that unlike the current study, this recent study divided the retinal capillaries into three layers in order to study the middle capillary plexus and also utilized different image processing methods to obtain the vessel density and GPD measurements.

Based on the International Council of Ophthalmology recommendations, eyes classified as referable in our study should technically be re-examined within 1 month,¹⁰ since we did not stratify the referable cohort by DR severity when calculating the ROC curves. Early detection and referral for treatment are crucial for preventing vision loss.^5,6 Recent advancements in the use of handheld color fundus photography, which can be incorporated into cell phones, show promise in reaching a wider community to screen for DR. These fundus cameras are currently cheaper, more convenient and capture a larger field of view compared to the images presented in the current study.⁴⁴ These devices coupled with artificial intelligence or computer-aided diagnostics have quickly gained traction for their ability to categorize color fundus photographs and screen for referral-warranted DR.⁸ On the other hand, the OCTA technology is gaining traction with increasing availability in clinics around the world. The evidence provided by the current study suggests OCTA can be used in the clinic by optometrists and ophthalmologists to provide quantitative data that correlates well with DR severity. This device can be used in addition to the standard ophthalmoscopic examination and color photography as a safety net for detecting eyes with severe DR. It can facilitate evidence-based decisions for requesting additional testing, such as fluorescein angiography, or for establishing follow-up intervals based on DR status. OCTA has many additional advantages over color photography, including depth-resolved, capillary level illustration of ischemia in DR as well as a detailed view of the severity DMI without the use of dye. A color coded map of GPD, which highlights the areas of deficit, may be a more practical visual tool than a vessel density map. These images may provide clues to resolve previously unexplained vision loss related to macular ischemia.^45,46 Furthermore, the OCTA device provides volumetric OCT data, which allows the user to reliably evaluate the structure of the retina on cross-section or en face. Volumetric OCT and OCTA data allow accurate assessment of DME as well as the integrity of the neuronal layers in relation to these capillary changes.⁴⁷ Compared to fundus photography, OCT is a more sensitive and reproducible measure of retinal thickening in eyes with DME.⁴⁸ One study found that of eyes with central-involved DME diagnosed on OCT, 26.9% to 32.7% lacked specific features such as hard exudates and therefore were missed using monocular fundus photographs.⁴⁹

Limitations of this study include a limited number of eyes in moderate (n = 16) and severe (n = 10) NPDR. Our study was also limited in the number of treatment naïve eyes with moderate NPDR or worse (n = 39) leading to a lower AUC for detecting clinically referable eyes for the treatment naïve subgroup. This limited the extent to which we could confidently assess the true sensitivity and specificity of vascular parameters in OCTA. Larger cohorts in future studies are needed to verify these results. This study did not obtain a full categorization of ETDRS stages by a reading center, which may have improved the accuracy of diagnoses. Another limitation is that the eyes with DME in the current study did not have a large amount of intraretinal fluid and further evaluation of GPD in eyes with DME is warranted. The definition of GPD as well as its 3-dimensional quantification also need further study since the theoretical 2D cutoff of 30 microns from the nearest blood vessel has yet to be validated in the living retina, as well as examined in a three-dimensional model.

In summary, we show that GPD is a useful parameter for detecting eyes with DR that could benefit from prompt referral or close follow-up with an ophthalmologist. We also found that the DCP layer performed the best at distinguishing these eyes, which was true for the entire cohort as well as the treatment naïve subgroup. We also established a threshold of 2.5% or greater for the DCP GPD as a first pass cut-off for eyes with referable DR. With this evidence that OCTA can distinguish eyes with referrable DR, we believe it has a role as an adjunct to conventional imaging methods for detecting referrable DR. Capitalizing on the many advantages of this technology, developing a color-coded map of perfusion deficits, using cross-sectional OCT data for assessing DME, and longitudinal follow-up metrics can provide great clinical value for evaluating DR and DMI.