Efficient hybrid state space architecture with linear attention enhances diabetic retinopathy severity assessment

Diabetic retinopathy (DR) grading from color fundus photographs remains challenging due to severe class imbalance in datasets and subtle inter-class differences within images. We propose DRMamba, a hybrid framework that couples four-directional Visual Selective Scan branches (VSS) with an Efficient Linear Attention Module (ELAM) and a TriBalance loss. Outperforming strong baselines by up to 1.6 pp in Accuracy (ACC) and 1.1 pp in Quadratic Weighted Kappa (QWK), DRMamba proves highly effective, registering ACC/QWK values of 0.847/0.851 on the DDR dataset and 0.905/0.938 on APTOS 2019. It is also efficient (23.8 M parameters; ~926 images/s on A6000 GPU). These results highlight DRMamba as an efficient and clinically relevant solution for DR grading, with promising applicability to broader medical tasks.