Toward Reliable Diabetic Retinopathy Severity Assessment via Cross-Domain Retinal Representation Learning

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, necessitating accurate and reliable automated severity grading systems for early diagnosis and timely intervention. However, existing deep learning methods often exhibit limited lesion localization, poor cross-dataset generalization, inadequate modeling of the ordinal progression of disease severity, and limited prediction reliability. To address these challenges, this study proposes a lesion-guided domain-adaptive framework that integrates an EfficientNetV2-S backbone for retinal feature extraction, lesion-guided attention, adversarial domain adaptation, Earth Mover's Distance-based ordinal learning, Monte Carlo (MC) Dropout for uncertainty estimation, and Grad-CAM for model interpretability. Experiments were conducted on a harmonized dataset created by combining the APTOS2019 and MESSIDOR datasets, with the original labels consolidated into three severity classes: No DR, Mild DR, and Severe DR. The proposed framework achieved an accuracy of 90.26%, precision of 90.10%, recall of 90.26%, F1-score of 90.14%, and a Quadratic Weighted Kappa (QWK) score of 89.28%. Uncertainty-aware evaluation using MC Dropout achieved 89.89% accuracy, 88.99% QWK, and a mean predictive entropy of 0.2571. Comparative evaluations, backbone comparisons, and ablation studies further demonstrated the effectiveness of the proposed components. These results indicate that the proposed framework provides accurate, robust, interpretable, and reliable DR severity assessment across heterogeneous retinal image datasets.