CS ‐ MSANet : Direction‐Aware Representation and Consistency‐Driven Feature Reconstruction for Medical Image Segmentation

Abstract

Medical image segmentation is crucial for computer‐aided diagnosis, yet accurately delineating complex anatomical structures remains challenging. Many existing architectures rely predominantly on isotropic local modeling or coupled channel–spatial attention, which can limit directional representation and flexible multiscale fusion. In addition, features transmitted through skip connections may exhibit semantic and spatial mismatch across encoder and decoder stages. To address these issues, CS‐MSANet adopts a CSWin‐based encoder for long‐range directional context modeling and introduces two dedicated refinement modules. In the decoder, the multiscale attention (MSA) block decouples channel and spatial reweighting and combines them with multishape convolutional branches to capture local patterns at complementary orientations and scales. The dynamic feature reconstructor (DFR) refines skip‐connection features through group cross‐attention, boundary enhancement, channel calibration, and dynamic fusion before encoder–decoder integration. Extensive experiments on Synapse, ACDC, and ISIC2018 demonstrate competitive or superior segmentation performance compared with existing approaches, while ablation and repeated‐run analyses further support the contributions and stability of the proposed design.