The Importance of Matching Optical Coherence Tomography Angiography Metrics to Diabetic Retinopathy Severity for Detecting Progression

Abstract

Purpose

The purpose of this study was to determine which macular optical coherence tomography angiography (OCTA) metric best captures early progression of capillary non-perfusion across diabetic retinopathy (DR) severities.

Methods

In this prospective, 1-year observational study, 208 patients with diabetes (320 eyes) underwent 3 × 3 mm macular OCTA at baseline, 6 months, and 12 months. Registered, averaged images yielded geometric perfusion deficits (GPDs), vessel density (VD), vessel length density (VLD) in the superficial and deep capillary plexuses (SCP and DCP), and foveal avascular zone area. Eyes were graded as non-referable or referable. Linear mixed models adjusted for age, sex, diabetes duration, hemoglobin A1c, and hypertension were conducted. Post hoc Dunnett's tests compared follow-ups with baseline within each severity group.

Results

A total of 709 eye visits were analyzed by OCTA. Across the cohort, referable DR, longer diabetes duration, and hypertension were independently associated with higher GPD values. In referable eyes, GPD-DCP increased at 6 months (P = 0.036) and 1 year (P = 0.016), whereas no other OCTA metric changed significantly at 6 months. In non-referable eyes, the only significant change was a decrease in VD-SCP at 1 year (P = 0.004).

Conclusions

Microvascular progression follows distinct, layer-specific patterns. In non-referable DR, capillary ischemia progresses more slowly and is more prominent in the SCP. In referable DR, GPD-DCP detected the earliest signs of microvascular progression and may serve as a promising biomarker, preceding vessel-based metrics. These findings suggest that the most informative OCTA metric should be tailored not only to the study question or timeline, but also to DR severity.

Diabetic retinopathy (DR) is a leading cause of visual impairment and blindness worldwide, with microvascular ischemia characterized by progressive capillary non-perfusion, contributing to visual decline.^1–3 Approximately one-third of patients diagnosed with diabetes develop DR,⁴ making it the primary cause of preventable blindness among working-age adults.

Based on their risk of progression to vision loss, eyes classified as referable DR require prompt evaluation by an ophthalmologist to ensure timely management and minimize the risk of irreversible vision loss.^5,6 Referable DR includes eyes with center-involving diabetic macular edema, or moderate non-proliferative DR (NPDR) or worse DR severity.⁷ A key component in managing these eyes is the assessment of retinal ischemia, which strongly influences prognosis and therapeutic decisions.⁸ In fact, the Diabetic Retinopathy Clinical Research Network (DRCR) identified baseline retinal nonperfusion and predominantly peripheral lesions on ultra-widefield fluorescein angiography (FA) as key risk factors for DR worsening.⁹ Moreover, we recently demonstrated that FA-based posterior retinal ischemia independently correlates with both best-corrected and low-luminance visual acuity in diabetes, underscoring the functional relevance of ischemia.¹⁰ With that in mind, considering FA is relatively invasive and quantitative metrics are not standardized, alternative quantitative methods to assess microvascular ischemia are highly desirable.¹¹

In that regard, optical coherence tomography angiography (OCTA) offers the distinct advantage of noninvasively visualizing perfused capillaries, offering depth-resolved quantitative metrics of the superficial and deep capillary layers. OCTA has revealed capillary information not otherwise seen with conventional FA and offers great promise in identifying eyes at risk of vision-threatening outcomes.¹² One of the limitations of the current OCTA software is the lack of agreement on the optimal metrics, and variability in thresholding methods used to quantify non-perfusion, which preclude wide-scale implementation.^13–15

To address this gap, Zhang et al. and Chen et al. introduced the concept of defining ischemia on OCTA based on a threshold distance of the retinal tissue from the nearest capillaries.^13,16 The geometric perfusion deficits (GPDs) use a threshold of 30 µm distance from the nearest capillary, considered the theoretical diffusion distance of oxygen, to define non-perfused, ischemic tissue.^13,17,18 Building on this concept, subsequent studies, including those from our group, demonstrated that GPD in the deep capillary plexus (DCP) correlates with angiographic nonperfusion on ultra-widefield FA, discriminates referable DR with high sensitivity and specificity, and even predicts vision-threatening outcomes in eyes with referable NPDR.^19–22 These findings highlight GPD as a promising, noninvasive marker of diabetic retinal ischemia, especially in the DCP.

Whereas cross-sectional evidence supports the association between GPD and DR severity, the longitudinal progression of GPD has not been fully characterized, nor have longitudinal changes in GPD been compared with other OCTA metrics — such as vessel density (VD), vessel length density (VLD), and foveal avascular zone (FAZ) area. Consequently, it remains uncertain which OCTA parameter is the most sensitive indicator of DR progression, particularly in longitudinal settings.

In light of these unresolved issues, our study adopts a prospective design, uses semi-automated quantification,²³ and leverages averaged OCTA images to enhance signal-to-noise ratio.^24,25 Through this refined approach, we aim to identify the most sensitive macular OCTA metric that tracks the progression of ischemia in DR. In this 1-year prospective study, we evaluated how the various OCTA metrics evolve over time, with particular emphasis on the difference between referable and non-referable DR. Our results have the potential to guide OCTA-based personalized management strategies for patients with DR and may provide potential endpoints for microvascular progression in clinical trials.

Methods

Study Subjects

This study is an interim analysis of a prospective, 2-year observational protocol that was approved by the Institutional Review Board of Northwestern University. We performed a longitudinal analysis of data from the first year, involving participants with diabetes mellitus (DM) who were enrolled between October 2021 and September 2023. All subjects provided written informed consent prior to enrollment, in compliance with the Health Insurance Portability and Accountability Act (HIPAA), and the study adhered to the tenets of the Declaration of Helsinki. Comprehensive ophthalmic examinations including slit-lamp microscopy, fundus examination, Early Treatment Diabetic Retinopathy Study (ETDRS) visual acuity testing, and spectral-domain optical coherence tomography (SD-OCT; Heidelberg Engineering, Heidelberg, Germany) were performed at baseline. Ultrawide-field FA and color photography using ultrawide-field scanning laser ophthalmoscopy (Optomap Panoramic 200, Optos PLC; Dunfermline, Scotland) were also performed.

Inclusion criteria consisted of adults (≥18 years) with either type 1 or type 2 DM but without fovea-involving diabetic macular edema (DME), defined by a central subfield thickness (CST) ≤300 µm.²⁶ Exclusion criteria included a history of intravitreal anti-VEGF or steroid treatment in the past 6 months, major ocular surgery in the previous 3 months or anticipated within the next 6 months, ocular conditions that might alter retinopathy status, and an HbA1c level exceeding 10%. Eyes with an OCTA quality score <6 or signal strength index <50 were also excluded. OCTA scans were acquired using RTVue-XR Avanti system (Optovue Inc., version 2017.1.0.151) utilizing split-spectrum amplitude-decorrelation angiography for angiographic data.²⁷

We initially planned to recruit approximately 200 participants covering approximately 260 study eyes across various DR severity levels: no clinically apparent DR (DM without DR), NPDR (mild, moderate, or severe), proliferative diabetic retinopathy (PDR) naïve, and PDR quiescent.²⁸ This target was based on pilot data from our group indicating correlation coefficients (r = 0.35–0.50) between baseline FAZ area and subsequent ischemic progression, leading us to calculate subgroup sample sizes via Fisher's z-transformation (2-sided α = 0.05, 80% power), with an additional 10% allowance for dropout. A total of 232 diabetic participants were recruited for this prospective study. For the present analysis, we used data collected between October 18, 2021, and October 1, 2024. We evaluated the results of OCTA analysis at the baseline, the 6-month follow-up, and the 1-year follow-up.

OCTA Imaging and Analysis

An overview of this entire image processing pipeline, including image averaging, is summarized in Figure 1. This workflow ensures a standardized approach for quantitative OCTA analysis.

Figure 1.

Overview of the Image Processing Pipeline for Optical Coherence Tomography Angiography Analysis. This figure provides an overview of the processing steps used for the quantitative analysis of OCTA images from the superficial capillary plexus (SCP), deep capillary plexus (DCP), and full retina slab. The workflow begins with the averaging of exported images for each slab to enhance vessel visualization and reduce noise. The foveal avascular zone (FAZ) and watermarks are first manually selected and then semi-automatically removed using a Fiji macro, which also calculates various OCTA metrics and saves the processed images for visualization. This integrated approach ensures a reliable assessment of retinal microvascular changes, allowing for a comprehensive evaluation of perfusion characteristics across different retinal layers.

First, we aimed to acquire at least five 3  ×  3 mm (304  ×  304 pixels) OCTA scans centered on the fovea with minimal motion artifacts for each eye. Using the instrument's default “Angio Retina” slab definitions, the retinal microvasculature was divided into the superficial capillary plexus (SCP), a combined intermediate-plus-DCP, and the full-retina slab. The SCP was defined from the internal limiting membrane to 10 µm above the inner plexiform layer (IPL), whereas the DCP slab extended from 10 µm above the IPL to 10 µm below the outer plexiform layer, thereby encompassing the intermediate capillary plexus. We did not analyze the intermediate capillary plexus separately because this two-plexus scheme is widely adopted and validated in DR research. The full retinal slab combines both SCP and DCP slabs, extending from the internal limiting membrane to 10 µm below the outer plexiform layer.

Only OCTA scans with a signal strength of 6 or greater were exported and registered using Fiji software (NIH, USA) to enhance image quality and improve signal-to-noise ratio.²⁴ The SCP angiogram with the highest signal intensity and minimal motion artifact was selected as the reference. Registration was performed using the “Register Virtual Stack Slices” plugin, where a rigid model was used for feature extraction, and an elastic registration model was applied to align the SCP slabs. The same transformation was applied to DCP and full retina slabs using transform virtual stack slices plugin. Finally, the registered images were averaged to enhance image quality and increase signal-to-noise ratio.

Subsequently, as previously described,¹⁹ we used a semi-automated macro on Fiji to measure the VD, VLD, and GPD in the averaged scans of each vascular layer. No additional spatial filtering was applied before thresholding because speckle noise was minimized by averaging; however, for construction of the large-vessel mask we first applied a 3 × 3 median (Despeckle) filter to remove isolated pixels, and then retained connected components with an area ≥30 pixel² (Analyze Particles, circularity 0–0.50). GPD was defined as any retinal region located more than 30 µm from the nearest blood vessel.¹³ To create the GPD maps, we first binarized the averaged angiograms using global thresholding with the Huang2 plugin for capillaries and the Max Entropy plugin for large vessels. Large- and small-vessel masks were therefore generated separately in the SCP, and all capillary-level metrics—including GPD—were derived exclusively from the capillary mask. In the SCP specifically, a 2-step approach was adopted: (1) the Huang2 thresholding with subsequent skeletonization to obtain the complete vascular network, and (2) subtraction of the large-vessel mask (generated by Max Entropy thresholding followed by a ≥30 pixel² size filter) so that only capillary segments contributed to the distance-transform–based GPD calculation. Because the transverse resolution of the OCTA system is too coarse to fully capture capillary dimensions, these binarized capillary maps were skeletonized, as described above. The VD was then quantified as the percentage of white pixels in the binarized images relative to the total area, whereas the VLD was calculated in the skeletonized images by summing the total length of skeletonized vessels and dividing by the total area. Next, a large-vessel mask was overlaid onto the skeletonized SCP angiogram prior to measuring the GPD. As noted above, the large-vessel mask was deliberately not applied to the DCP because Max Entropy fails to remove projection artifacts in this slab, and doing so would falsely eliminate genuine capillary-free regions. Areas corresponding to large vessels were then subtracted from both the GPD and total area measurements to avoid distortion from projection or shadowing artifacts. Finally, the FAZ was delineated manually on the full retinal slab using ImageJ's polygon selection tool; both the FAZ and the watermark region were excluded from the GPD calculations. The intra-grader reliability and reproducibility were verified in a previous study.²³

Representative examples of each image processing step, including binarization, skeletonization, and GPD measurement, are shown in Figure 2. These examples provide a visual reference for VD, VLD, and GPD calculations in the SCP and DCP.

Figure 2.

Processing and Quantification of Optical Coherence Tomography Angiography Images of the Superficial and Deep Capillary Plexus: Vessel Density, Vessel Length Density, and Geometric Perfusion Deficits. (A, B) Averaged OCTA images of the superficial capillary plexus (SCP) and deep capillary plexus (DCP), respectively. (C, D) Binarized images of the SCP and DCP layers obtained using the Huang thresholding method, which converts the grayscale OCTA images into binary images where vessels appear as white areas and the background as black. These images are used to calculate vessel density by measuring the proportion of vessel area relative to the total image area. (E, F) Skeletonized images of the SCP and DCP layers, where vessel structures are reduced to single-pixel-wide lines while preserving their original shape. This processing allows for the quantification of vessel length density by measuring the total vessel length. (G, H) Images of the SCP and DCP layers processed to identify areas that are at least 30 µm away from the nearest vessel (red areas). These images are used to calculate geometric perfusion deficits by analyzing the perfusion-free areas relative to the total area, excluding the foveal avascular zone (FAZ). In the DCP layer, an additional step is performed to subtract large vessels before the final measurements.

Data Collection

Data were collected across several critical domains to comprehensively assess each participant. Demographic and general health information included age, sex, and systemic comorbidities, including a history of hypertension, ischemic heart disease, renal dysfunction, and cerebrovascular disease, which were collected as binary categorical variables at baseline. Diabetes-specific data encompassed the type of diabetes (type 1 or type 2), duration of diabetes, and the most recent hemoglobin A1c (HbA1c) level, serving as an indicator of glycemic control. Ocular health assessments were conducted for each eye, documenting lens status (whether cataract surgery had been performed), axial length measurements obtained using the IOL Master device (Carl Zeiss Meditec, Jena, Germany), and best corrected visual acuity at baseline and 1 year after baseline. Additionally, the number of OCTA scans and the OCTA quality score were recorded for each eye at baseline, as well as at the 6-month follow-up and the 1-year follow-up. DR severity was evaluated by two independent graders masked to the OCTA metrics using ultrawide-field FA or color photography, following the International Clinical Diabetic Retinopathy severity scale as previously outlined.^19,28 Any disagreements between the two graders were resolved by a third grader. Eyes with DM but no apparent DR (DM without DR) and mild NPDR were defined as non-referable eyes, whereas eyes with moderate NPDR or more severe DR were defined as referable eyes.²⁹

Statistics

Statistical analyses were conducted using R software (version 4.4.3; R Foundation for Statistical Computing, Vienna, Austria) and RStudio (version 2024.04.2; Posit, Boston, MA, USA). Linear mixed models were used to investigate the main effects and interaction effects of time (baseline, the 6-month follow-up, or the 1-year follow-up) and DR severity (non-referable eyes versus referable eyes) on OCTA metrics. Age, sex, DM duration, HbA1c, and a history of hypertension were included as covariates to control for their potential confounding effect.³⁰ Both linear (first degree) and quadratic (second degree) terms of time were included to capture nonlinear changes over time. A time-by-DR severity interaction term (Time × DR severity) was included in the model to assess whether the effect of time on each outcome measure differed between the non-referable and referable DR groups. Patient and eye (nested within patient) were modeled as random effects to account for repeated measurements within individuals and eyes. All linear mixed models were fitted using the restricted maximum likelihood method.

Post hoc comparisons were conducted using Dunnett's test to compare each time point (the 6-month follow-up and the 1-year follow-up) against baseline within the non-referable and the referable DR subgroups, as well as in the entire cohort. Estimated marginal means (EMMs) were calculated to provide adjusted means for each group and time point, accounting for covariates and missing data under the assumption of missing at random. These EMMs offer a more reliable representation of trends compared with unadjusted means. To assess multicollinearity, generalized variance inflation factors (GVIFs) were calculated and adjusted by raising them to the power of 1/(2*Df), as recommended for models with categorical variables and interaction terms.³¹ Residual diagnostics, including Q-Q plots and residual plots, were used to evaluate normality and homoscedasticity.

As a sensitivity analysis, the same linear-mixed-model specification was fitted separately within the non-referable and referable cohorts.

To assess whether dropout rates differed across disease severity groups, a chi-square test was performed at both the 6-month and 1-year follow-ups. Similarly, to evaluate potential bias due to dropout, we compared medical and sociodemographic characteristics between participants with and without follow-up visits at both the 6-month and 1-year time points using the Mann-Whitney U test and chi-square test, as appropriate.

A P value of < 0.05 was considered statistically significant.

Results

Of the 232 diabetic participants considered for this study, a total of 208 patients (320 eyes) met the eligibility criteria and were included in the final analysis. The median (interquartile range [IQR]) of age was 60.5 (IQR = 52–68) years, and 107 participants were women. Type 2 DM was present in 168 patients, with a median DM duration of 18.4 (IQR = 8.8–27.4) years and a median HbA1c of 7.2 (IQR = 6.5–8.0; Table 1). Table 2 summarizes the baseline ocular characteristics. Of 320 eyes, 112 were classified as non-referable DR eyes, whereas 208 were classified as referable DR eyes. At baseline, among the 112 patients with data from both eyes, 8 patients exhibited asymmetrical referable status, with one eye classified as non-referable and the fellow eye as referable (Supplementary Table S1). Among the 320 study eyes, 192 (60.0%) and 197 (61.6%) completed the 6-month and 1-year follow-ups, respectively. A total of 144 eyes contributed data at both the 6-month and 12-month time points. The chi-square test showed no significant difference in dropout based on DR severity (6 months: P = 0.481 and 1 year: P = 0.346) or between referable and non-referable groups (6 months: P = 0.304 and 1 year: P = 0.727), indicating that loss to follow-up was likely non-differential (Supplementary Table S2). Similarly, medical and sociodemographic characteristics were also largely comparable between participants with and without follow-up visits (Supplementary Table S3). The average number of OCTA scans and the Q-score per eye per visit were 4.86 ± 1.23 and 7.53 ± 1.43, respectively (mean ± standard deviation). During the 1-year study period, 8 eyes (8 patients) received anti-VEGF injections, and 15 eyes (13 patients) underwent panretinal photocoagulation.

Table 1.

Baseline Characteristics of 208 Study Patients

Variable	Value
Age, y	60.5 (52–68)
Sex, F/M	107/101
DM type (1/2)	40/168
DM duration, y	18.4 (8.8–27.4)
HbA1c (%)	7.2 (6.5–8.0)
Hypertension, n (%)	138 (66)
Ischemic heart disease, n (%)	41 (20)
Renal dysfunction, n (%)	43 (21)
Cerebrovascular disease, n (%)	13 (6)

DM, diabetes mellitus.

Data were presented as median (interquartile range) or N (%).

Table 2.

Baseline Ocular Characteristics of 320 Study Eyes

Variable	Value
BCVA, letters	84 (79–88)
Axial length, mm	23.60 (22.87–24.54)
DR severity, n (%)
DM no DR	74 (23)
Mild NPDR	38 (12)
Moderate NPDR	63 (20)
Severe NPDR	41 (13)
PDR naïve	23 (7)
PDR quiescent	81 (25)
Lens status, n
Phakia	227 (71)
Pseudophakia	93 (29)

BCVA, best-corrected visual acuity; DR, diabetic retinopathy; DM, diabetes mellitus; NPDR, non-proliferative DR; PDR, proliferative DR.

Data were presented as median (interquartile range) or N (%).

Table 3 presents the results of the linear mixed models for GPD-SCP and GPD-DCP across the entire study cohort. Referable DR, DM duration, and hypertension were identified as independently significant explanatory variables for both GPD-SCP and GPD-DCP, after controlling for DR severity, DM duration, HbA1c, hypertension, age, and sex. Specifically, referable DR showed the strongest association with both GPD-SCP (β = 4.607, P < 0.001) and GPD-DCP (β = 2.652, P < 0.001), indicating significant impact of DR severity on GPD in both capillary plexuses. Regarding systemic factors, DM duration was significantly associated with GPD-SCP (β = 0.134, P < 0.001) and GPD-DCP (β = 0.038, P = 0.046), suggesting that longer disease duration is associated with higher perfusion deficits. Similarly, hypertension was found to be a significant predictor for both GPD-SCP (β = 2.064, P = 0.026) and GPD-DCP (β = 1.706, P = 0.001). On the other hand, over the 1-year period, follow-up time did not have a statistically significant effect on GPD in the overall cohort. The adjusted GVIF values ranged from 1.014 to 1.691, showing low multicollinearity. As a sensitivity analysis, we fitted the same linear mixed model separately within the non-referable and referable cohorts; the results are presented in Supplementary Tables S4 and S5, respectively. These analyses confirmed that the key associations observed in the overall cohort were generally consistent within each DR severity subgroup. The results of the linear mixed models for other OCTA metrics in the entire study cohort are shown in Supplementary Tables S6–S8. These tables highlight the influence of systemic factors, referable DR status, and follow-up duration (time) on VD, VLD, and FAZ area.

Table 3.

Results of Linear Mixed Models Assessing Factors Associated With Geometric Perfusion Deficits in Superficial and Deep Capillary Plexuses

	GPD-SCP					GPD-DCP
Fixed Effects	Estimate	SE	df	T Value	P Value	Estimate	SE	df	T Value	P Value
(Intercept)	7.059	3.152	189.1	2.239	0.027*	3.511	1.754	184.2	2.001	0.047*
Time (linear)	0.679	0.352	381.5	1.930	0.054	0.193	0.218	421.5	0.888	0.375
Time (quadratic)	−0.171	0.363	377.8	−0.471	0.638	0.054	0.225	416.4	0.239	0.811
DR severity = Referable	4.607	0.822	240.2	5.604	<0.001 *	2.652	0.432	251.8	6.141	<0.001 *
Age	−0.050	0.034	184.2	−1.480	0.141	−0.024	0.019	177.4	−1.248	0.214
Sex	1.095	0.821	184.8	1.333	0.184	0.434	0.458	178.3	0.947	0.345
DM duration	0.134	0.034	211.3	3.934	<0.001 *	0.038	0.019	203.4	2.010	0.046 *
HbA1c	−0.193	0.357	203.8	−0.542	0.588	−0.291	0.198	198.9	−1.473	0.142
Hypertension	2.064	0.920	184.9	2.244	0.026 *	1.706	0.512	178.7	3.331	0.001 *
Time (linear) × referable	0.286	0.432	384.3	0.663	0.508	0.222	0.267	424.1	0.829	0.408
Time (quadratic) × referable	0.508	0.458	379.2	1.108	0.269	−0.253	0.284	418.2	−0.893	0.372

df, degrees of freedom; DM, diabetes mellitus; DR, diabetic retinopathy; GPD-DCP, geometric perfusion deficits in deep capillary plexus; GPD-SCP, geometric perfusion deficits in superficial capillary plexus; SE, standard errors.

Linear mixed models were used. Time, DR severity, age, sex, DM duration, HbA1c, and hypertension were adjusted. To account for intra-patient and intra-eye correlations, both patient and eye (nested within patient) were modeled as random effects. The model included both linear and quadratic terms of time to account for potential nonlinear changes over 1 year.

^*P < 0.05.

The P values in bold represent statistical significance.

Post hoc Dunnett's tests from linear mixed models were used to evaluate OCTA metrics in eyes with referable and non-referable DR (Table 4). At 6 months, referable DR eyes showed a statistically significant increase in GPD-DCP compared to baseline (5.82 ± 0.30 to 6.36 ± 0.33, P = 0.036, presented as estimated marginal means ± standard errors), but no differences were seen in the non-referable eyes. At the 1-year follow-up, referable eyes showed significant changes in several metrics including GPD-SCP (13.04 ± 0.54 to 14.41 ± 0.58, P < 0.001), GPD-DCP (5.82 ± 0.30 to 6.41 ± 0.32, P = 0.016), VLD-SCP (14.03 ± 0.20 to 13.55 ± 0.22, P = 0.008), and VLD-DCP 17.73 ± 0.16 to 17.36 ± 0.17, P = 0.006). On the other hand, in non-referable eyes, only VD-SCP showed a statistically significant decline at 1 year (45.00 ± 0.50 to 43.50 ± 0.70, P = 0.004). Table 4 also includes results from the entire cohort to confirm consistency of longitudinal trends across the study population.

Table 4.

Longitudinal Evaluation of OCTA Metrics Using Estimated Marginal Means With Post Hoc Dunnett's Tests

	Ocular Severity	Baseline	6 Mo	P Value†	1 Y	P Value§
GPD-SCP, %	Non-referable	8.43 ± 0.71	9.12 ± 0.76	0.273	9.39 ± 0.76	0.101
	Referable	13.04 ± 0.54	13.31 ± 0.58	0.678	14.41 ± 0.58	<0.001 *
	Entire	10.74 ± 0.46	11.22 ± 0.49	0.208	11.90 ± 0.49	<0.001 *
GPD-DCP, %	Non-referable	3.43 ± 0.39	3.50 ± 0.42	0.949	3.70 ± 0.42	0.577
	Referable	5.82 ± 0.30	6.36 ± 0.33	0.036 *	6.41 ± 0.32	0.016 *
	Entire	4.63 ± 0.26	4.93 ± 0.28	0.193	5.06 ± 0.28	0.045 *
VD-SCP, %	Non-referable	45.00 ± 0.50	44.75 ± 0.70	0.804	43.50 ± 0.70	0.004 *
	Referable	41.06 ± 0.54	40.65 ± 0.54	0.408	40.33 ± 0.54	0.061
	Entire	43.03 ± 0.43	42.70 ± 0.46	0.429	41.92 ± 0.46	<0.001 *
VD-DCP, %	Non-referable	47.90 ± 0.49	48.23 ± 0.55	0.728	48.14 ± 0.56	0.849
	Referable	45.33 ± 0.36	45.47 ± 0.42	0.896	45.24 ± 0.41	0.944
	Entire	46.62 ± 0.30	46.85 ± 0.34	0.674	46.69 ± 0.35	0.954
VLD-SCP, mm−1	Non-referable	15.82 ± 0.27	15.94 ± 0.30	0.780	15.31 ± 0.30	0.057
	Referable	14.03 ± 0.20	13.81 ± 0.23	0.334	13.55 ± 0.22	0.008 *
	Entire	14.92 ± 0.17	14.88 ± 0.19	0.912	14.43± 0.19	0.001 *
VLD-DCP, mm−1	Non-referable	19.39 ± 0.21	19.63 ± 0.23	0.287	19.40 ± 0.23	0.992
	Referable	17.73 ± 0.16	17.64 ± 0.17	0.693	17.36 ± 0.17	0.006 *
	Entire	18.56 ± 0.13	18.63 ± 0.15	0.714	18.38 ± 0.15	0.169
FAZ area, mm²	Non-referable	0.312 ± 0.023	0.311 ± 0.023	0.994	0.311 ± 0.023	0.996
	Referable	0.399 ± 0.017	0.413 ± 0.017	0.102	0.413 ± 0.017	0.066
	Entire	0.355 ± 0.014	0.362 ± 0.015	0.435	0.362 ± 0.015	0.389

^†Comparison between 6 mo and baseline follow-ups.

^§Comparison between 1-y and baseline follow-ups using Dunnett's multiple comparison test.

^*P < 0.05.

Data were presented as estimated marginal means ± standard errors. All analyses were performed at the eye level, with intra-patient correlation accounted for by including patient as a random effect in the linear mixed models used to generate the estimated marginal means.

The P values in bold represent statistical significance.

Discussion

In this longitudinal study, we demonstrated that eyes with referable DR displayed statistically significant progression of capillary non-perfusion in both the SCP and DCP at 1 year. Notably, GPD-DCP showed statistically significant changes as early as 6 months, underscoring its potential as an early indicator of microvascular progression in referable DR. In contrast, among non-referable eyes, significant progression of non-perfusion was seen in the SCP only (VD-SCP) at 1 year. These findings indicate that non-perfusion progresses at different patterns in the two macular plexuses with advancing DR severity, which has important implications for designing personalized monitoring strategies.

Our results are generally consistent with previous longitudinal studies using OCTA^32–35; however, differences in patient characteristics and analysis protocols are important to highlight here. First, Aschauer et al. conducted a 2-year prospective observational study in patients with type 2 diabetes, of whom 90% had diabetes without retinopathy.³² They reported significant progression in VD-SCP, with no difference in the VD-DCP, consistent with our results in non-referable eyes. Similarly, Wang et al. followed DM without DR for 1 year and reported that VD-SCP declined in eyes that developed DR, whereas VD-DCP remained stable.³³

Focusing on eyes with stable laser-treated proliferative DR, Thottarath et al. performed a 1-year longitudinal analysis of OCTA and demonstrated a significant increase in FAZ metrics, whereas VD in the SCP and DCP remained stable.³⁴ Interestingly, our referable DR cohort showed no significant progression in the FAZ area or the VD metrics over 1 year. We believe the differences in FAZ findings can be attributed to the difference in DR stage (more advanced in the former study) and the use of image averaging in our study. Averaging enhances the signal-to-noise ratio and improves the visualization of the FAZ, potentially leading to more robust analysis compared to studies that rely on single-frames.²⁵ Notably, a recent study in advanced NPDR (ETDRS Diabetic Retinopathy Severity Scale [DRSS] 43 and 47) found that DCP skeletonized VD (comparable to VLD), progressed significantly over 2 years, whereas the corresponding SCP metrics did not, the particular vulnerability of deeper vessels in referable DR stages.³⁵ Compared to their study, our cohort encompassed a broader range of DR severity (approximately DRSS 43 to 65), including more advanced stages, revealing changes in the DCP as early as 6 months, which suggests that DCP impairment may be a consistent and rapidly progressive feature of eyes with referable DR. Furthermore, we found that in referable DR, not only the GPD-DCP but also the VLD-DCP showed significant progression over 1 year. This suggests that progression predominantly impacted smaller capillaries, leading to detectable changes in VLD, which quantified skeletonized vessel length. In contrast, the VD metric used in the Thottarath study³⁴ is based on vessel area and may over-represent larger caliber vasculature, and thus underestimate changes at the smaller capillaries. In aggregate, these findings emphasize the utility of metrics like VLD and GPD, which are sensitive to capillary-level changes, for monitoring progression over short spans in high-risk eyes.^36,37

To complement these longitudinal findings, we additionally performed a supplemental baseline analysis to assess the diagnostic performance of GPD-DCP. This parameter independently differentiated referable from non-referable eyes with excellent accuracy (area under the curve [AUC] = 0.909, 95% confidence interval [CI] = 0.876–0.943), in line with our prior cross-sectional findings.²⁰

In referable DR, our findings showed that GPD-DCP detected ischemic progression earlier than VLD-DCP or any metrics at the SCP, suggesting heightened vulnerability in the deep plexus and the utility of using the “distance to nearest capillary” as threshold for defining non-perfusion. Anatomically, the DCP is situated closer to the metabolically demanding photoreceptors, making it especially prone to the harmful effects of reactive oxygen species.^38–40 Moreover, the stratified organization of the retinal vasculature renders the DCP inherently more susceptible to compromised perfusion, and its predominantly venous outflow may further predispose it to ischemic insults.^41–43 These structural and physiological factors collectively exacerbate microvascular compromise in the DCP, potentially explaining more rapid changes at the GPD-DCP in referable DR.

In addition to these subgroup-specific findings, our analysis of the entire cohort demonstrated that systemic factors significantly influence OCTA metrics. Notably, GPD in both the SCP and DCP was independently associated with DR severity, hypertension, and longer diabetes duration, highlighting the cumulative impact of these factors on macular perfusion deficits. Similarly, VD and VLD in both the SCP and DCP were significantly associated with at least two of the following variables—follow-up time (e.g. baseline, 6 months, and 12 months), DR severity, hypertension, and diabetes duration—indicating that microvascular rarefaction captured by OCTA is influenced by factors related to diabetes as well as overall vascular health (see Supplementary Tables S6–S8). Interestingly, the FAZ area was the only metric that showed a significant association with sex, in addition to DR severity, with female participants exhibiting larger FAZ areas. This finding aligns with previous reports that identify sex-based differences in FAZ size in healthy individuals, and further underscores the importance of considering baseline patient characteristics and strategies that adjust for the baseline, pre-pathologic FAZ size to facilitate accurate estimation of the impact of microvascular disease.^44–46

This study has several limitations. Because DR was classified based on baseline imaging, we could not evaluate the progression of DR severity at the 6-month and 1-year follow-ups. To address this, we included an interaction term between time and baseline DR severity, to account for their potential impact on the trajectory of microvascular changes over time. We divided the study eyes into two groups based on referable DR status, but we acknowledge this approach may have concealed important differences within the DR sub-stages. A finer classification could provide deeper insights into microvascular progression patterns, but it would inevitably reduce subgroup sample sizes, thus lowering the statistical power. Consequently, we opted for a two-group classification to strike a balance between clinical relevance and the ability to detect meaningful differences. Prospective longitudinal studies with larger cohorts incorporating DRSS-based stratification will be required to capture the full gradient of microvascular progression across DR severities. We adjusted our analysis for age to partially mitigate for age-related differences in microvascular susceptibility but cannot definitively rule out these concerns. We acknowledge that the precision of diabetes duration may differ between type 1 and type 2 diabetes, as the exact onset of type 2 diabetes is often uncertain. Our cohort included a relatively higher proportion of patients with type 1 diabetes compared with the general diabetic population. These factors may limit the generalizability of our findings to the broader population, which is predominantly composed of individuals with type 2 diabetes. To minimize the impact of DME on OCTA image quality and ischemia quantification, we excluded eyes with center-involved DME at baseline. Whereas necessary for methodological rigor, this exclusion may have biased the cohort toward eyes with more stable retinal phenotypes. We acknowledge that our findings may not be directly generalizable to eyes with coexisting DME. Additionally, a relatively high proportion of patients (38.9% and 36.5% at the 6-month and 12-month follow-ups, respectively) did not complete follow-up, which could have impacted our ability to detect changes in OCTA metrics. To address this, we examined dropout rates across severity groups at follow-ups (see Supplementary Table S2), finding no difference with respect to disease severity. In addition, our use of linear mixed models under the missing-at-random (MAR) assumption provides robustness against data loss. Therefore, although the reduced sample size may have limited our statistical power, it is less likely to have introduced systematic bias between the referable and non-referable DR groups. Furthermore, our study cohort did not include healthy eyes without diabetes, so the false positive rate of the GPD algorithm in detecting ischemia in healthy eyes could not be directly assessed. This remains an important area for future validation, ideally in studies including a dedicated healthy control cohort. Moreover, because GPD was calculated from 2D en face projections of the SCP and DCP slabs, depth information was not incorporated; therefore, inter-plexus overlap in three dimensions could not be assessed, and ischemic areas may have been underestimated.³⁷

We also want to highlight notable strengths of our study. The prospective design with interval assessments of OCTA at 6 and 12 months, along with the relatively large cohort spanning different stages of DR, lends robust internal validity and ensures a broad representation of disease severity. By using linear mixed models that adjust for key confounders, the study effectively addresses the relationship between DR severity and OCTA metrics. Importantly, we incorporated averaged OCTA images and GPD—a physiologically grounded marker of capillary ischemia beyond the oxygen diffusion limit—enabling us to detect subtle microvascular damage that might be missed on traditional OCTA parameters. Finally, our standardized imaging protocol using averaged images and semi-automated quantification further bolster reproducibility, with the potential for these metrics to be integrated into future clinical trials and disease-monitoring algorithms for DR.

In conclusion, our longitudinal analysis suggests that non-perfusion in the DCP may be a particularly sensitive biomarker of progressive microvascular compromise in referable DR, revealing changes as early as 6 months. In non-referable eyes, alterations in VD within the superficial plexus emerged over 1 year, indicating that microvascular parameters may progress at different patterns in the capillary plexuses with advancing DR. These findings underscore the potential value of incorporating metrics such as the GPD or VLD, which prioritize small vessel changes, in DR monitoring especially among eyes with referable DR.^19,21 Our findings may also have potential implications for clinical trials and clinical practice. Because GPD-DCP was the earliest OCTA metric to show progression in referable DR, it may help identify eyes that require closer surveillance. With further confirmation in larger cohorts, GPD-DCP could potentially serve as an objective exploratory endpoint in trials, and it could be instrumental in monitoring microvascular progression in referable DR. Future studies with longer follow-up periods and more granular DR severity stratifications could further validate the utility of these metrics and refine individualized clinical risk assessments. Overall, our study highlights the importance of using different OCTA metrics based on DR severity, supporting a more personalized approach to monitoring microvascular changes in DR.