Differential Capillary and Large Vessel Analysis Improves OCTA Classification of Diabetic Retinopathy

Abstract

Purpose: This study aimed to investigate the impact of distinctive capillary–large vessel (CLV) analysis in optical coherence tomography angiography (OCTA) on the classification performance of diabetic retinopathy (DR).

Methods: This multicenter study analyzed 212 OCTA images from 146 patients, including 28 controls, 36 diabetic patients without DR (NoDR), 31 with mild non-proliferative DR (NPDR), 28 with moderate NPDR, and 23 with severe NPDR. Quantitative features were derived from the whole image as well as the parafovea and perifovea regions. A support vector machine classifier was employed for DR classification. The accuracy and area under the receiver operating characteristic curve were used to evaluate the classification performance, utilizing features derived from the whole image and specific regions, both before and after CLV analysis.

Results: Differential CLV analysis significantly improved OCTA classification of DR. In binary classifications, accuracy improved by 11.81%, rising from 77.45% to 89.26%, when utilizing whole image features. For multiclass classifications, accuracy increased by 7.55%, from 78.68% to 86.23%. Incorporating features from the whole image, parafovea, and perifovea further improved binary classification accuracy from 83.07% to 93.80%, and multiclass accuracy from 82.64% to 87.92%.

Conclusions: This study demonstrated that feature changes in capillaries are more sensitive during DR progression, and CLV analysis can significantly improve DR classification performance by extracting features that are specific to large vessels and capillaries in OCTA. Incorporating regional features further improves DR classification accuracy. Differential CLV analysis promises better disease screening, diagnosis, and treatment outcome assessment.

Early disease diagnosis and objective treatment assessment require a quantitative analysis of retinal neurovascular changes. Various diseases, such as diabetic retinopathy (DR), strokes (cerebrovascular accidents), hypertensive retinopathy, and different vascular disorders, can lead to alterations and irregularities in the retinal vasculature.^1,2 Notably, these diseases impact capillaries and pre-/post-capillary vessels (i.e., large vessels) differently due to their inherent structural differences. The distinctive changes observed in capillaries and large vessels during disease progression emphasize the significance of employing differential capillary–large vessel (CLV) analysis in various medical conditions, including diabetes, hypertension, stroke, cardiovascular disease, and other vascular disorders.^3–5 The incorporation of CLV analysis capabilities into clinical imaging devices can enhance the precision of disease detection and classification. Nevertheless, the limited resolution and contrast of fundus images present challenges in detecting retinal microvascular abnormalities, particularly when assessing smaller capillary-level blood vessels around the fovea.^6,7 

As a modality extension of optical coherence tomography (OCT), OCT angiography (OCTA) has proven its effectiveness in non-invasively detecting microvascular changes down to the capillary level that are associated with ocular diseases.⁸ A substantial body of research has concentrated on the quantitative analysis of OCTA images to quantify microvascular changes and objectively identify and categorize retinal diseases.^9–24 They have demonstrated the efficacy of OCTA quantitative analysis in detecting microvascular changes, offering a nuanced approach to disease categorization and early detection. Furthermore, some studies have directed their focus toward distinguishing between arteries and veins in OCT and OCTA imaging to improve the detection and classification of retinal diseases.^25–32 

Despite the recognized importance of OCTA analysis in DR classification, there has been limited focus on the separate evaluation of capillaries and large vessels. In a study involving 3 × 3-mm OCTA scans, Lei et al.³³ employed a multiscale line detector to identify large vessels and computed some quantitative features for both large vessels and capillaries. Their findings revealed a trend of large vessel enlargement and a significant reduction in superficial capillaries with increasing DR severity. Subsequently, in a similar study involving 6 × 6-mm OCTA scans,³⁴ they calculated vessel density and skeleton density for large vessels and capillaries. Their results confirmed an enhancement of large vessels in DR, accompanied by decreased capillary densities. Recently, Cheng et al.³⁵ introduced various image processing techniques for 3 × 3-mm OCTA images to explore the large vessels and capillaries separately, aiming to assess the correlations between OCTA parameters and visual acuity in epiretinal membrane. These studies collectively underscore the potential of differential CLV analysis in OCTA for enhancing the sensitivity of DR classification. 

Our study aimed to enhance the classification of DR by applying a novel differential CLV analysis using OCTA. The rationale behind this objective is the hypothesis that differential CLV analysis can improve OCTA sensitivity to reveal subtle vascular changes due to DR progression, which traditional OCTA analysis may overlook. In this study, we employed a deep learning–based approach to segment both large vessels and capillaries within OCTA images. This segmentation enabled the computation of various quantitative features for both large vessels and capillaries, facilitating a deeper analysis. Following an assessment of the quantitative feature changes during the progression of DR, we leveraged the features both before and after CLV segmentation for the binary and multiclass classification of DR stages. The results of these classifications provided valuable insights into the importance of the quantitative features and the impact of differential CLV analysis on the performance of the classification models. 

Methods

Data Acquisition

This research obtained ethical approval from the Institutional Review Board (IRB) at the University of Illinois Chicago (UIC; protocol #2016-0752) and adhered to the tenets of the Declaration of Helsinki. En face OCTA images (6 × 6-mm scans) were collected between February 2017 and April 2023 at both UIC and the National Taiwan University Hospital (NTUH). The dataset encompassed a total of 212 OCTA images, distributed as follows: 52 control, 48 diabetic patients without DR (NoDR), 37 with mild non-proliferative DR (NPDR), 39 with moderate NPDR, and 36 with severe NPDR scans. A summary of participant demographics and diabetes-related parameters is presented in Table 1. Control subjects and diabetic patients without and with varying stages of DR were recruited from the UIC retina clinic and NTUH. Inclusion criteria encompassed individuals 18 years or older who were diagnosed with type II diabetes mellitus. Exclusion criteria included individuals with macular edema, proliferative DR (PDR), prior vitrectomy surgery, a history of ocular disorders other than DR, the presence of cataracts or minor refractive errors, and ungradable or low-quality OCT images. Three board-certified retina specialists (YH, MJH, JIL) classified patients as NoDR or into a specific NPDR stage based on the Early Treatment Diabetic Retinopathy Study (ETDRS) criteria based on clinical findings derived from fundus photography as the gold standard for diagnosing DR. All patients underwent a comprehensive anterior and dilated posterior segment examination. Control OCTA images were obtained from healthy volunteers who provided informed consent for OCT/OCTA imaging. Deidentified diabetic datasets were acquired for retrospective analysis. The IRB waived the requirement for patient informed consent, but the study adhered to patient privacy and confidentiality guidelines outlined by the IRB. 

Table 1.

 

Demographics of the Healthy Subjects and Patients With NoDR or NPDR

 

Spectral-domain (SD) en face OCTA data were obtained using AngioVue SD-OCT devices (Optovue, Fremont, CA, USA) at both the UIC retina clinic and NTUH. The OCT devices operated at a rate of 70,000 A-scans per second, providing an axial resolution of approximately 5 µm and a lateral resolution of about 15 µm for 6 × 6-mm scans. For this study, only superficial OCTA images segmented at the level of the retinal nerve fiber layer and the ganglion cell layer were utilized. After image reconstruction, deidentified en face OCTA images were extracted from the ReVue software interface (Optovue) for subsequent processing. 

Feature Extraction

Based on our previous publication,³¹ a deep-learning network, MF-AV-Net, with an overall accuracy of 96.02%, was employed to segment large arteries and veins within OCTA images. By separating the areas occupied by large vessels from those of the capillaries, both large vessel and capillary maps were generated. Figures 1A, 1D, and 1G show a representative OCTA image of a total vasculature map, a large vessel map extracted from MF-AV-Net, and a capillary map, respectively. To enhance the visualization of vasculature, a Hessian-based multiscale Frangi filter¹⁹ was applied, with the outcomes binarized for clarity, as shown in Figures 1B, 1E, and 1H. The binarized images were then skeletonized to remove boundary pixels while maintaining the integrity of the vasculature structures,²¹ as demonstrated in Figures 1C, 1F, and 1I. In Figure 1C, the fovea (diameter 1 mm), parafovea (diameter 1–3 mm), and perifovea (diameter 3–6 mm) are highlighted with red, orange, and green circles, respectively. It should be mentioned that the OCTA layer indicator area, marked by a green rectangle in Figure 1C, bottom left, was consistently excluded from all images. 

Figure 1.

 

Illustrating feature extraction from an OCTA image. (A) OCTA image of total vasculature map. (B) Binarized total vasculature image. (C) Skeletonized total vasculature image illustrating fovea, parafovea, and perifovea areas. (D) Large vessel map extracted from MF-AV-Net. (E) Binarized large vessel image. (F) Skeletonized large vessel image. (G) Capillary map. (H) Binarized capillary image. (I) Skeletonized capillary image.

From total vasculature, large vessel, and capillary maps, we extracted five quantitative features: perfusion intensity density (PID), blood vessel density (BVD), vessel skeleton density (VSD), blood vessel caliber (BVC), and vessel area flux (VAF). The detailed procedure for PID calculation is outlined in our recent publication.³² For BVD, also known as vessel area density (VAD), which is the ratio of the image area occupied by blood vessels,⁸ we employed a fixed threshold binarization strategy.¹⁶ VAF was calculated using the procedure described by Abdolahi et al.²⁰ Utilizing the binarized and skeletonized images presented in Figure 1, we followed the procedures described by Yao et al.⁸ to compute BVC and VSD. 

Utilizing total vascular images enables the computation of features for the total vascular system, encompassing both large vessels and capillaries. Upon CLV segmentation, quantitative features for large vessels and capillaries can be derived from the respective large vessel and capillary maps. To facilitate clarity, we use notations such as T-PID, L-PID, and C-PID to denote PID in the total vasculature images, large vessel map, and capillary map, respectively. Additionally, we define various ratios between the features, such as the large vessel–capillary ratio, large vessel–total vasculature ratio, and capillary–total vasculature ratio for each feature. For example, the large vessel–capillary PID ratio is the division of L-PID by C-PID; the large vessel–total PID ratio is the division of L-PID by T-PID; and the capillary–total PID ratio is the division of C-PID by T-PID. Quantitative features for the total vasculature (i.e., T-PID, T-BVD, T-VSD, T-BVC, and T-VAF) can be calculated without distinguishing between large vessels and capillaries. Furthermore, these quantitative features can be computed across various regions, including the whole image, parafovea, and perifovea, within distinct cohorts. 

Statistical Analyses

For statistical analysis of quantitative features, each eye was considered a unique observation for subjects with images of both eyes. We employed χ² tests to evaluate the distribution of sex and hypertension among various groups. To assess the normality of age and diabetes duration distribution in different groups as presented with mean and standard deviation in Table 1, we conducted the Shapiro–Wilk test. Comparisons of age and diabetes duration distribution were carried out using analysis of variance (ANOVA). In all comparisons, statistical significance was determined with a P < 0.05. 

Classification Model

To prepare for this study, we evaluated the performance of various classifiers, including decision trees, random forest, Gaussian naïve Bayes, logistic regression, and support vector machines (SVMs). The selected model, an SVM classifier equipped with radial basis function (RBF) kernels, demonstrated the best performance with hyperparameters set to C = 1 and gamma = scale. SVMs are favored for their efficiency in handling high-dimensional spaces, where the number of features can exceed the number of samples, making them well suited for small to medium-sized datasets. Given our high-dimensional data, limited classes, and small dataset size, the SVM classifier emerged as the optimal choice. We utilized SVM classifiers for binary classification across different groups, applying a one-versus-one strategy to evaluate the impact of CLV analysis on binary classification performance. Furthermore, we implemented multiclass SVM classification using a one-versus-rest strategy, to assess the effect of CLV analysis on multiclass classification outcomes. Both binary and multiclass classifications were subjected to a 10-fold cross-validation to ensure the robustness of our retrospective study. The performance of the SVM classifiers was assessed using two primary metrics: accuracy and the area under the receiver operating characteristic (ROC) curve (AUC). For all classification tasks, we adopted sequential forward selection (SFS),³⁶ a method based on the greedy search algorithm, to identify the optimal subset of features that maximize classification accuracy. The SFS method begins with no features and iteratively adds the feature that most significantly enhances classification accuracy, continuing this process until no further improvements can be observed. 

Results

CLV Features

Table 1 provides an overview of demographic characteristics of the dataset used in this study. The statistical analysis revealed that the distributions of age and diabetes duration were normal. Furthermore, the analysis indicated no significant differences between groups concerning age and sex (P = 0.1063, ANOVA; P = 0.0844, χ² test). Additionally, there was no variation in the prevalence of hypertension across diabetes groups (P = 0.3865, χ² test). However, among the diabetic groups, a significant distinction was observed in terms of diabetes duration (P = 0.0217, ANOVA). PID, BVD, VSD, BVC, and VAF, as well as their corresponding ratios, were computed for total vasculature, large vessels, and capillaries in the whole image. Employing all available features for classification and using the SFS algorithm to identify the optimal feature subset that maximizes classification accuracy can yield robust classification performance across diverse groups. 

Binary Classifications

We utilized the SFS algorithm alongside SVM classifiers to identify the optimal subset of features for binary classification among various groups. We extracted features from different image regions for classification purposes. Features from the whole image were categorized under “whole image features,” and those from the entire image, along with parafoveal and perifoveal regions, were classified as “features in three regions.” Figure 1C depicts a representative image with labels for these regions. Initially, the classifier used features solely from the total vasculature, including T-PID, T-BVD, T-VSD, T-BVC, and T-VAF. Following CLV analysis, we incorporated additional features from large vessels and capillaries, including their ratios, into the classifier. 

Table 2 showcases a side-by-side comparison of the accuracy and AUCs for various binary classifications both before and after CLV analysis, using features from total vasculature, large vessels, capillaries, and their ratios in the whole image and in three regions. In Table 2, we simplify the groupings for clarity: ModSev represents a combined group of moderate and severe cases, and NPDR encompasses all non-proliferative diabetic retinopathy stages, including mild, moderate, and severe. The impact of CLV analysis on classification performance is evident. When utilizing features from the whole image, we observed a substantial increase in mean accuracy, rising by 11.81% (from 77.45% to 89.26%). Similarly, incorporating features from the three regions resulted in an accuracy enhancement of 10.73% (from 83.07% to 93.80%). This finding underlines the effectiveness of CLV analysis in enhancing the accuracy of DR classifications using OCTA images. 

Table 2.

 

Performance Comparison of Binary Classifications Before and After CLV Analysis

 

Multiclass Classifications

We employed the SFS algorithm in conjunction with multiclass SVM classifiers using the one-versus-rest strategy. A comparison of accuracy and AUCs for multiclass classifications was carried out, both before and after the implementation of CLV analysis, incorporating features from total vasculature, large vessels, capillaries, and their ratios. These comparisons were made for different regions: the whole image and the three regions, as outlined in Table 3. The CLV analysis significantly enhanced the multiclass classification of DR. Specifically, accuracy improved by 7.55% (from 78.68% to 86.23%) with the analysis of whole image features. When expanding the feature analysis to the three regions, accuracy increased by 5.28% (from 82.64% to 87.92%). The ROC curves with Youden index values in Figure 2 present the performance of the one-versus-rest approach across various groups, along with the micro-average curve. This visual representation and the increase in AUCs and Youden index values effectively illustrate the enhanced performance of the multiclass classifier after applying CLV analysis. 

Table 3.

 

Performance Comparison of Multiclass Classifications Before and After CLV Analysis Using All Features in Different Regions

 

Figure 2.

 

ROC curves with Youden index values for multiclass classification using data from the whole image before CLV analysis (A), from the whole image after CLV analysis (B), from three regions before CLV analysis (C), and from three regions after CLV analysis (D).

Discussion

Diabetes manifests differently in the retinal capillaries and large vessels, with DR primarily affecting microvasculature. DR is characterized by damage to the capillaries, leading to clinical exam findings such as microaneurysms, hemorrhages, and cotton-wool spots. Pericytes are essential for retinal capillary structure and function, and it has been well established that pericyte loss is an early feature of DR.³⁷ Shen et al.³⁸ reported that, during the early stage of DR, the density of the superficial retinal capillary plexus reduces significantly in all areas. As the disease progresses, it may also affect large vessels, leading to symptoms such as beading and potentially causing vascular occlusion. In advanced stages of DR, OCTA revealed venous beading and intraretinal microvascular abnormalities.⁴ This underscores the importance of using differential CLV analysis to differentiate between capillaries and large vessels in evaluating DR and possibly other retinal conditions impacting capillaries. Integrating CLV analysis into clinical imaging devices has the potential to improve disease detection and classification accuracy. The advent of OCTA offers a noninvasive method to identify microvascular distortions in ocular diseases with capillary-level resolution. 

This study aimed to evaluate the effectiveness of CLV analysis in the binary and multiclass classification of DR stages and the effect of DR on large and capillary vessels. To our knowledge, this is the first study to apply differential CLV analysis in OCTA for DR classification. This study introduces a novel technique specifically designed to segregate capillaries and large vessels within OCTA images, facilitating comprehensive CLV analysis. We computed five quantitative OCTA features, including PID, BVD, VSD, BVC, and VAF, for total vasculature, large vessels, and capillaries, along with their corresponding ratios, across different regions. These ratios enable us to quantify and analyze the relative feature changes in total vasculature, large vessels, and capillaries. 

Tables 2 and 3 show that features derived from large vessels tend to reduce the performance of both binary and multiclass classification models compared to those derived from the total vasculature. In contrast, features from capillaries improve the classification performance. Additionally, incorporating features from the total vasculature, large vessels, capillaries, and their corresponding ratios obtained through CLV analysis significantly enhances classification performance. This is consistent with prior research indicating that DR predominantly targets capillaries.^37,38 

For binary classifications, CLV analysis yields average accuracy improvements of 11.81% and 10.73% using features from the whole image and the three regions, respectively. The mean AUCs also increased by 9.43% and 8.33% for features from the whole image and the three regions, respectively. In multiclass classifications, CLV analysis resulted in accuracy increases of 7.55% and 5.28%, along with AUC improvements of 11.38% and 8.97%, for features from the whole image and the three regions, respectively. These findings highlight the importance of CLV analysis in extracting features specific to large vessels or capillaries, effectively differentiating between the stages of DR. The incorporation of features from three regions improves classification accuracies for both binary and multiclass DR classifications. The findings from other studies^39,40 directly correlate with our results showing improved DR classification accuracy when incorporating three-region features derived from OCTA images. This improvement suggests that the spatial distribution of vascular changes within parafovea and perifovea holds valuable diagnostic information that can enhance the model performance. Our observations also indicate that DR classification using only features from the capillary network yields results closely comparable to using all features after differential CLV analysis. Therefore, focusing on capillary features after differential CLV analysis could simplify and enhance the diagnostic process, underscoring the critical role of the capillary network in early DR detection. 

Although our multicenter study enhances the external validity of our findings across diverse populations and clinical settings, the limited dataset size is a limitation of this study; therefore, for some subjects, both eyes were included in the statistical analysis of quantitative features. There may be some correlation between right and left eyes due to genetics and environmental factors. Our findings lay the groundwork for future research, especially in exploring the application of differential CLV analysis in other retinal diseases. Our current study offers robust, interpretable, and clinically relevant results, but future work could explore the integration of CLV analysis with deep learning techniques to automate DR detection and classification. This study is based on a retrospective analysis for proof-of-concept demonstration. Planning subsequent prospective studies would be beneficial to verify the potential of integrating CLV analysis with various clinical instruments for clinical deployment. 

In conclusion, this study effectively showcased the utility of CLV analysis in OCTA for both multiclass and binary classification of DR stages. It suggests the potential of leveraging differential CLV analysis in OCTA for improved disease diagnosis and classification. The implementation of CLV analysis notably improved classification performance by extracting features related to large vessels and capillaries. Regional analysis further improves classification accuracy, compared to whole image analysis. These findings provide valuable insights into the potential integration of CLV analysis into clinical applications, aiming to improve sensitivity for effective disease diagnosis, treatment, and management. 

Acknowledgments

Supported by grants from the National Eye Institute (P30 EY001792, R01 EY023522, R01 EY030101, R01EY029673, R01EY030842), Research to Prevent Blindness, and the Richard and Loan Hill Endowment. 

Disclosure: M. Abtahi, None; D. Le, None; B. Ebrahimi, None; A.K. Dadzie, None; M. Rahimi, None; Y.-T. Hsieh, None; M.J. Heiferman, None; J.I. Lim, None; X. Yao, None