MCEPANet: a connectivity-edge guided attention network for robust medical image segmentation with multi-scale boundary preservation.

Medical image segmentation is an important technique for analyzing medical imaging data and supporting clinical decision-making. However, existing methods still face challenges in boundary delineation due to non-rigid anatomical deformation caused by respiratory motion and organ displacement, as well as the loss of fine details during hierarchical feature extraction. To address these issues, we propose a multi-scale connectivity-edge pixel-level attention network (MCEPANet). Specifically, we introduce a statistical pixel-level connectivity extraction algorithm to characterize anatomical topology through spatial correlation analysis, thereby improving robustness to organ deformation and inter-subject variability. We further design a connectivity-edge pixel-level attention (CEPA) module to integrate edge features derived from traditional edge detection with learned pixel-level connectivity priors via channel-wise attention. CEPA is embedded at multiple network stages to capture edge information at different scales and improve boundary localization. Experimental results on the Synapse multi-organ abdominal segmentation dataset show that MCEPANet achieves an average DSC score of 84.15% and an average 95% Hausdorff Distance (HD95) of 14.88 mm, outperforming existing state-of-the-art two dimensional segmentation methods. Cross-domain validations on the ACDC cardiac dataset and the ISIC 2018 skin lesion dataset confirm its generalizability, highlighting its versatility for different clinical segmentation tasks.