AttMamba: Integrating Dual-Attention and Vision Mamba for Diabetic Retinopathy Grading

Diabetic retinopathy (DR) is the leading cause of blindness among the global working population. It is crucial to early screening and timely diagnosis in mitigating vision deterioration and preventing permanent blindness. The varying sizes and complex structures of different types of DR lesions present challenges for accurate grading. However, existing models suffer from long training time and insufficient extraction of small lesion features, which makes it difficult to learn the subtle differences between adjacent grades of DR images. This paper proposes the AttMamba framework for automatic DR grading, which integrates an attention mechanism with the state space model to improve training efficiency and the ability to extract small lesion features. The efficient 2D scanning module is integrated with VMamba to scan the whole image with fixed step sizes to accelerate the computation speed. Moreover, the squeeze-and-excitation module and global context module are proposed to locate regions of interest and extract channel attention and global attention features. It can learn the fine-grained differences between different grades of DR images for precise grading. Experiments are conducted on three public and one personal datasets. The accuracy of the proposed method achieved 0.832/0.789/0.756/0.889 on APTOS 2019, DDR, FGADR, and RUDR datasets, which improved by 7.4%/9.6%/17.2%/8.5% in accuracy compared to VMamba. The proposed model provides an effective solution for early diagnosis and timely treatment of diabetic retinopathy.