CSWin-MDKDNet: cross-shaped window network with multi-dimensional fusion and knowledge distillation for medical image segmentation

Sci Rep. 2026 Mar 2;16(1):11532. doi: 10.1038/s41598-026-40690-5.

Abstract

In recent years, deep learning has achieved significant advancements in medical image segmentation. Medical image segmentation is fundamental to computer-aided diagnosis, yet challenges persist in balancing local detail preservation and global context modeling. This paper proposes CSWin-MDKDNet, a novel Transformer-based architecture enhanced with Multi-dimensional Selective Fusion (MDSF) and Knowledge Distillation Loss (KD-loss). The MDSF module refines multi-scale feature fusion through channel-spatial attention, while KD-loss mitigates feature redundancy in deep layers. Evaluated on the Synapse (multi-organ CT), ACDC(cardiac MRI) and ISIC2018 datasets, our model achieves state-of-the-art performance, with 81.82% DSC (Synapse), 91.76% DSC (ACDC) and 91.64% DSC(ISIC2018), outperforming existing methods in accuracy.

Keywords: Cross-shaped window attention; Deep learning; Knowledge distillation; Medical image segmentation; Multi-dimensional feature fusion.

MeSH terms

  • Algorithms
  • Deep Learning*
  • Humans
  • Image Interpretation, Computer-Assisted* / methods
  • Image Processing, Computer-Assisted* / methods
  • Magnetic Resonance Imaging / methods
  • Neural Networks, Computer
  • Tomography, X-Ray Computed