Deformable Encoder–Decoder Network for Accurate Retinal Vessel Segmentation

ABSTRACT Retinal vessel segmentation represents a pivotal technique in precision medicine. To address the limitations of existing methods in capturing irregular vessel morphology and effectively fusing multi‐scale features, this study proposes a lightweight deformable encoder–decoder network, termed ResDC‐Net, for retinal vessel segmentation. The network comprises three core components: an encoder, skip‐connection modules, and a decoder. Both the encoder and decoder are constructed primarily using Residual Deformable Convolution (ResDC) modules. These modules employ deformable convolutions to dynamically adapt to vessel geometry and enhance feature flow via residual connections, effectively overcoming the representational limitations of conventional convolutions on curved vessels and bifurcation structures, thereby improving the modeling capability for vessel morphological variations. A Group Multi‐axis Hadamard Product Attention (GHPA) module is introduced as the skip‐connection module, which integrates multilevel features from the encoder outputs and enhances the complementary relationship between semantic information and spatial details, thereby facilitating the recovery of vascular boundary particulars. Experiments are conducted on three public datasets—STARE, CHASE_DB1, and HRF—achieving accuracies of 97.43%, 97.62%, and 97.18%, respectively. The results demonstrate the efficacy, lightweight design, and superior performance of the proposed network in the task of retinal vessel segmentation.