Abstract

Info

Author(s):: QU Xiaobo1; YU Su2; 1. College of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China 2. Engineering Training Center, Shanghai University of Engineering Science, Shanghai 201620, China

Keywords:: Keywords: medical image segmentation deep learning channel enhancement residual network spatial attention network dynamic loss function

Abstract:: Abstract: An improved retinal blood vessel segmentation algorithm is proposed to address the problem that U-Net algorithm can not segment the tiny peripheral blood vessels and deal with noise interference in fundus image segmentation. The proposed method introduces channel enhancement residual network into U-Net algorithm to optimize the U-Net architecture and make the network recognize more retinal microvessels, uses spatial attention network to eliminate noise and better highlight blood vessels, and replaces the fixed weight of U-Net algorithm with dynamic weight in the calculation of the loss function for enabling the neural network to learn a robust feature map. The experiment on DRIVE dataset show that the improved algorithm exhibits better performance of accuracy and sensitivity, which are 2.12% and 7.51% higher than the original U-Net algorithm, and 1.20% and 2.55% higher than DCU-Net algorithm.