SE-CA Swin-Unet: Dual attention enhanced transformer for CT image segmentation

Abstract

This paper proposes the SE-CA Swin-Unet, a dual-attention enhanced U-shaped Transformer architecture for CT image segmentation. Instead of simply stacking existing attention components, we systematically investigate the optimal cascading order of Squeeze-and-Excitation (SE) and Coordinate Attention (CA) in skip connections. The SE module strengthens channel-wise dependencies, while the CA mechanism captures precise spatial correlations, and their optimized ordering yields improved segmentation accuracy. Evaluated on the Synapse multi-organ CT dataset, our method achieves a mean Dice Similarity Coefficient (DSC) of 80.25 ± 0.51% and a 95th percentile Hausdorff Distance (HD95) of 24.38 ± 1.00 mm. Ablation studies confirm the effectiveness of combining SE and CA, offering a promising framework for precise clinical organ segmentation.