MSPDD-net: Mamba semantic perception dual decoding network for retinal image vessel segmentation.
Journal:
Computers in biology and medicine
Published Date:
May 19, 2025
Abstract
In the Retinal Image Vessel (RIV) segmentation task, due to existing a large number of low-contrast capillaries in the image usually leads to the problem of poor segmentation accuracy. To address this issue, this study aims to fully model the global contextual information of the image and mine edge features to further improve the accuracy of RIV segmentation. We propose a novel Mamba Semantic Perception Dual Decoding Network (MSPDD-Net). Firstly, a dual-encoding path incorporating Position Sensitive Cross-Layer Interactive (PSCLI) strategy was designed, which first uses convolution to extract the initial encoding features of RIV images, and then constructs the second encoding path using PSCLI mechanism to obtain refined encoding features containing more semantic information. Then, a novel Mamba Full-Scale Semantic Perception (M-FSSP) module was designed, which was embedded into the bottleneck layer of the network to capture the full-scale semantic feature of RIV image, and learn more discriminative semantic representations to guide decoding. Finally, a dual-decoding path combining Full-Scale Semantic Injection (FSSI) and Wavelet Edge Attention (WEA) was designed, which first uses the FSSI module for the first decoding to alleviate semantic gaps, and then uses the WEA mechanism for secondary fusion decoding of initial decoding features to highlight edge features, and improve vessel segmentation accuracy. Comparative experimental results based on three publicly available retinal image datasets, i.e., DRIVE, STARE, and CHASEDB1, show that the segmentation accuracies of the designed MSPDD-Net are 97.45 %, 97.76 %, and 97.88 %, respectively. Among them, the STARE dataset showed a 1.59 % improvement compared to the baseline network, indicating that the proposed MSPDD-Net provides an effective solution for retinal image vessel segmentation.