Abstract

Info

Title:: Deformable encoder-decoder network for accurate retinal vessel segmentation

Author(s):: YUE Danming1; WANG Xuelin2; DENG Shenmi1; RUAN Yongshuo1; JIANG Jing1; 1. Smart City College, Beijing Union University, Beijing 100101, China 2. School of Mechanical and Electrical Engineering, Qingdao City University, Qingdao 266106, China

Keywords:: medical image retinal vessel segmentation residual deformable convolution group aggregation bridge lightweight network

Abstract:: To overcome the drawbacks of existing approaches in capturing irregular vascular morphologies and integrating multi-scale features, this study proposes a lightweight deformable encoder-decoder network named ResDC-Net for retinal vessel segmentation. The network consists of 3 components: an encoder, skip connection modules, and a decoder. Both the encoder and decoder are mainly composed of residual deformable convolution modules. These modules utilize deformable convolutions to dynamically adapt to vascular morphology and enhance feature propagation through residual connections, which effectively compensates for the limited representation capacity of traditional convolutions for curved vessels and branches, thereby improving the modeling for vascular deformations. A group aggregation bridge module serves as the skip connection module. This module integrates multi-depth information output by the encoder and the mask information generated by the decoder at each stage, strengthening the complementarity between semantic information and spatial details, which is conducive to restoring vessel edge details. Experiments on the public STARE, CHASE_DB1, and HRF (resolution: 876×584, 1 752×1 168) datasets show that the proposed method achieves specificity of 98.66%, 98.67%, 98.43% and 98.42%, and accuracy of 97.47%, 97.63%, 97.18% and 97.17%, respectively, demonstrating its effectiveness, lightweight property, and superiority for retinal vessel segmentation.