MRAMnet: a multi-region attention network with contrastive disentangled learning for diabetic retinopathy grading

Diabetic retinopathy (DR) is a leading cause of preventable blindness, requiring accurate severity grading for timely clinical intervention. Current automated grading methods face two key challenges: inadequate capture of regional lesion correlations, and poor performance on underrepresented severe grades due to class imbalance. This study developed and validated MRAMnet, a novel deep learning framework addressing these limitations through multi-region attention mechanisms and contrastive disentangled learning. Using two public fundus image datasets — APTOS2019 (3,662 images) and EYEPACS (35,126 images) — MRAMnet partitions fundus images into quadrants, processed through a dual-branch architecture combining channel-spatial attention and transformer mechanisms. A Decoupled Supervised Contrastive Loss (DSCL) function balances feature learning across all severity grades. Evaluated via 5-fold cross-validation, MRAMnet achieved classification accuracy of 89.07% (APTOS2019) and 90.31% (EYEPACS), representing a 3–5% improvement over state-of-the-art methods. For severe grades, F1-score improvements of 8.7% (severe non-proliferative DR) and 6.5% (proliferative DR) were observed. MRAMnet demonstrates strong potential clinical utility in supporting DR screening programmes, particularly for sight-threatening cases requiring urgent referral.