Retinal Vessel Segmentation Based on Attention Mechanism and Multi-Path U-Net

To solve the problem that the existing algorithms cannot accurately segment the end of tiny blood vessels and the segmentation results are easily affected by optical contrast and the diseased area,a retinal vessel segmentation algorithm is proposed,which combining attention mechanism and multi-path U-Net. Firstly, a dual-path U-Net is designed to extract coarse and fine-grained blood vessels separately through texture and structure branches. The semantic guidance module is designed to fully integrate the deep and shallow features. Secondly, a residual block introducing attention mechanism and DropBlock is designed to replace the ordinary convolution block, which improves the segmentation effect of blood vessels in complex background area and prevents over-fitting. Finally, the output image of the dual-path U-Net and the original image are put into the feature refinement module for feature extraction and fusion, which further refines the blood vessel segmentation results. The accuracy of the proposed algorithm on DRIVE, STARE and CHASEDB1 datasets are 97.01%, 96.43%, and 97.52%; the sensitivity are 80.31%, 84.38% and 81.61%; and the AUC(Area Under Curve) are 98.67%, 98.06%, and 98.83%, respectively. The experimental results show that the comprehensive segmentation performance of the proposed algorithm is better than other algorithms.