Multi-scale Attention Net for Retina Blood Vessel Segmentation

U-Net has been recently employed to solve the medical image segmentation problems such as retina blood vessel segmentation. The structure with skip connections in U-Net can effectively concatenate the features of encoding and decoding branches. In this paper, we propose a Multi-scale Attention Net (MA-Net) based on U-Net where the variation of scale inside the paths and relevant regions in feature maps are considered. There are several advantages of the proposed architecture for segmentation tasks. First, multi-scale connections inside encoding and decoding paths can fuse information within and between blocks at the same time and ensure better feature representation. Second, MA-Net integrates attention mechanism to suppress irrelevant regions in feature maps and highlight salient features. Experimental results illustrate that the proposed method outperforms the compared algorithms on segmentation tasks for blood vessel.