Global-local collaborative framework integrating visual Mamba and Kolmogorov-Arnold networks for robust diabetic retinopathy grading

Automatic grading of diabetic retinal disease requires modeling global retinal structure and fine-grained local lesion properties together. Most approaches are limited by restricted receptivity fields of convolutional networks or large computational load of Vision Transformers. To solve this problem, we propose MIL-Kamba, a global–local dual branch collaborative learning framework for DR Grading. In the global branch, pyramid encoders are built by visual Mamba with KAN (Kolmogorov–Arnold network) to model long-range dependency with linear complexity and also improve nonlinear representation capability. In the local branch, we introduce task adaptive multiple instance learning mechanism in which critical features are aggregated by adaptive-temperature attention. Multiclass classification, ordinal regression and branch-consistency constraints are used to capture progressive and ordered nature of DR Grading. Experimental results on the APTOS 2019 and DDR datasets show that MIL-Kamba achieves QWK scores of 92.8% and 86.2%, respectively. With 8.77 M parameters and 3.3G FLOPs, the model facilitates 92 FPS inference, addressing the dual requirements of diagnostic precision and computational efficiency for real-world clinical screening.