MSU-Net

Retinal vessel segmentation is widely used in the diagnosis of eye diseases, and the effect of segmentation plays a crucial role in whether doctors can correctly diagnose diseases. To further improve the accuracy of the automatic segmentation method, a network structure named Multi-Scale U-Net (MSU-Net) based on deep learning is proposed in this paper. The network combines Atrous Spatial Pyramid Pooling (ASPP) module to extract multi-scale information, making the U-Net more suitable for segmentation of complex and changeable vessel structures. We evaluate the network on two public databases, DRIVE and STARE. The Accuracy (ACC), Sensitivity (SEN), Specificity (SPE) and Dice coefficient on the DRIVE database are 0.9667, 0.8159, 0.9805 and 0.8059, respectively. These indicators are respectively 0.9732, 0.8272, 0.9866 and 0.8400 on the STARE database. Experiments show that the network has excellent segmentation results, and has state-of-the-art performance indicators on the STARE database, which fully proves the outstanding performance of the network.