Interpretable Machine Learning-Based Concentric Regional Analysis of OCTA Images for Enhanced Diabetic Retinopathy Detection

Abstract

Diabetic retinopathy (DR) remains a major cause of vision loss in patients with diabetes, and earlier recognition of retinal vascular abnormalities may improve risk stratification and clinical follow-up. Optical coherence tomography angiography (OCTA) provides a noninvasive way to visualize the retinal microvasculature and may detect DR-related changes before they are evident on routine clinical assessment. In this work, we investigated whether dividing OCTA images into anatomically defined retinal regions could improve DR classification and clarify which regions carry the greatest discriminative information. The study included 188 OCTA images: 67 from normal eyes, 57 from eyes with mild DR, and 64 from eyes with moderate DR. Each image was divided into seven concentric regions centered on the fovea, and vessel-density features were extracted from each region. Ten machine learning classifiers were trained and compared at the regional level. For each region, the best-performing classifier was retained, and the final prediction was obtained with a majority-voting ensemble. To examine model behavior, Local Interpretable Model-Agnostic Explanations (LIME) were applied. Performance was also compared with that of a transfer-learning MobileNet model trained on whole OCTA images. On the held-out patient-level test set, the ensemble model achieved 97% accuracy, 98% precision, 97% recall, and a 97% F1-score for three-class classification. These results were higher than those obtained with the tested whole-image transfer-learning baselines. The interpretability analysis consistently identified the parafoveal regions as the most informative for classification. Among the seven regions, Region 3 showed the highest overall contribution, followed by Regions 2 and 5, whereas Region 5 became more influential in moderate DR. These results suggest that regional analysis of OCTA-derived vessel density can improve both classification performance and interpretability in DR assessment. The findings also indicate that parafoveal vascular alterations carry substantial discriminative value in distinguishing normal, mild DR, and moderate DR cases. Validation in larger, independent cohorts from multiple centers will be necessary to confirm the generalizability of these findings.