MSA-Net: MultiScale Attention Network for Diabetic Retinopathy Lesion Segmentation

Diabetic Retinopathy (DR) is an ophthalmic complication caused by diabetes, leading to retinal vascular damage and is an important cause of blindness in patients. Deep learning-based lesion segmentation has shown great potential in assisting physicians in diagnosis and early prevention, but it is challenging to precisely segment DR lesions due to the large number of types, varying sizes, and intra-class variations of DR lesions, which occupy a small area in the fundus map. This paper proposes a novel lesion segmentation network to enhance segmentation accuracy. First, a proposed Multiscale Feature Fusion Block (MSFF) is applied in encoder to integrate and enrich multiscale features. Then, the proposed Dilated Attention Block (DAB) is used to optimize model's ability to extract global and local features, focusing on key information. Finally, the proposed Local-Global Attention Block (LGAB) is introduced in skip connections to effectively fuse deep and shallow features, enhancing the amount of information in the decoder while improving feature selectivity. We conducted ablation experiments on the IDRiD dataset to validate the effectiveness of each component and analyze them in comparison with existing methods. Experimental results indicate that the proposed method achieves superior performance in DR lesion segmentation.