MCEPANet: A connectivity-edge guided attention network for robust medical image segmentation with multi-scale boundary preservation
- 11 June 2026
- journal article
- Published by IOP Publishing in Biomedical Physics & Engineering Express
- Vol. 12 (4) , 045020
Abstract
Medical image segmentation serves as a critical technology to link medical imaging data with clinical decision-making. However, existing segmentation methods still face significant challenges in boundary delineation: the spatial inconsistency caused by non-rigid deformations of anatomical structures induced by respiratory motion and organ shifts, and the loss of fine details during hierarchical feature extraction, leading to degraded edge precision.To address these issues, we propose a novel Multi-scale Connectivity-Edge Pixel-level Attention Network (MCEPANet). Specifically, we introduce a statistical pixel-level connectivity extraction algorithm that formally characterizes anatomical topology through spatial correlation analysis, improving the model's robustness against organ deformation and inter-subject variability. A Connectivity-Edge Pixel-level Attention (CEPA) module is proposed to adaptively integrate geometric edge features extracted from traditional edge detection technique with the learned pixel-level connectivity priors through channel-wise attention mechanisms. This design explicitly mitigates edge information degradation and significantly improves boundary representation capabilities. The MCEPANet incorporates the CEPA module at multiple network depths to capture edge details across different spatial scales, alleviating the boundary blurring problem prevalent in medical image segmentation tasks.Extensive experiment results on the Synapse multi-organ abdominal segmentation dataset demonstrate MCEPANet's superiority, achieving an average Dice Similarity Coefficient (DSC) of 84.15% and an average 95% Hausdorff Distance (HD95) of 14.88 mm, outperforming the existing state-of-the-art two dimensional segmentation methods. Cross-domain validation on the ACDC cardiac dataset confirms exceptional generalizability, highlighting its versatility for different clinical segmentation tasks.Keywords
Funding Information
- National Natural Science Foundation of China (61902282)
- National Natural Science Foundation of China (62071330)
This publication has 20 references indexed in Scilit:
- TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformersMedical Image Analysis, 2024
- EGD-Net: Edge-guided and differential attention network for surface defect detectionJournal of Industrial Information Integration, 2022
- Enhance Connectivity of Promising Regions for Sampling-Based Path PlanningIEEE Transactions on Automation Science and Engineering, 2022
- Edge-aware U-net with gated convolution for retinal vessel segmentationBiomedical Signal Processing and Control, 2022
- U-Net Transformer: Self and Cross Attention for Medical Image Segmentation2021
- BiconNet: An edge-preserved connectivity-based approach for salient object detectionPattern Recognition, 2021
- nnU-Net: a self-configuring method for deep learning-based biomedical image segmentationNature Methods, 2020
- Attention gated networks: Learning to leverage salient regions in medical imagesMedical Image Analysis, 2019
- Intracranial Vessel Wall Segmentation Using Convolutional Neural NetworksIEEE Transactions on Biomedical Engineering, 2019
- The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesionsScientific Data, 2018