Retinal vessel segmentation method based on improved U-Net

Abstract

Blood vessels are the main anatomical structure of the fundus retina. Retinal blood vessel segmentation images have been widely used in the judgment of cardiovascular and cerebrovascular diseases and retinal diseases. Therefore, appropriate fundus retinal blood vessel segmentation method is of great significance for the detection of retinal diseases. Based on U-Net, the original convolution structure in the encoding part is replaced by the Res-Se module, and the CBAM module is introduced in the skip connection part to achieve fine-grained feature fusion, thereby improving the network's ability to segment the subtle features of retinal vessels. Experiments on the CHASEDB1 dataset show that the proposed model has certain improvements in accuracy, sensitivity, and specificity indicators. This model can more accurately segment retinal vessels and demonstrate better segmentation performance.