Abstract

Info

Title:: Application of residual U-Net combined with three-space attention in retinal vessel segmentation

Author(s):: HANG Yiliu1; ZHANG Qiong1; QIU Jianlin1; 2; YANG Yuwei1; 1. School of Computer and Information Engineering, Nantong Institute of Technology, Nantong 226000, China 2. School of Information Science and Technology, Nantong University, Nantong 226000, China

Keywords:: Keywords: retinal vessel deep learning multi-level residual three-space attention U-Net

Abstract:: To addresses the issues of low contrast and inaccurate segmentation of tiny vessels in retinal images, a U-shaped network incorporating multi-level residuals and three-space attention mechanism is proposed. In encoding stage, a multi-level residual module is added after inputting original images for preserving image features, and additionally, batch normalization and Dropout are integrated into the residual module to prevent vanishing gradient and feature data redundancy within the deep network. In decoding stage, a three-space attention mechanism is adopted to assign different weights to the features from the original images, down-sampled images, and up-sampled images, thus enhancing feature texture and position information, and achieving precise segmentation of tiny blood vessels. Experimental results on a public color fundus image dataset demonstrate that the proposed algorithm achieves higher accuracy (0.985), specificity (0.991), sensitivity (0.829), and AUC (0.985) than the existing algorithms. Moreover, the vessel maps obtained by the comparison with the gold standard are of significant reference value in clinic.