Multiscale GAN-based fundus vascular segmentation algorithm

This paper addresses the problem of vascular segmentation of retinal fundus images, which is prone to breakage and missing microvessel segmentation, from the perspective of enabling the network to acquire more abundant and significant semantic information. First, using U-Net as the backbone network, the novel residual attention module RSCA and multiscale transformer as generators, and the binary classification model with dense jump connections as discriminators, which are combined to form a multiscale GAN network. Second, setting the joint loss function for parameter update. The accuracy, sensitivity, specificity, and AUC of the constructed model on the DRIVE dataset reached 98.57%, 84.13%, 98.63%, and 97.80%, respectively, which is at the level of human experts.