CGMU-Net: A Classification-Guided Cross-Interaction Network for Computer-Assisted Dermoscopy Skin Lesion Segmentation
Abstract Accurate and effective segmentation of dermoscopic skin lesions is critical for clinical diagnosis. Despite recent advances in deep learning, challenges such as irregular shapes, blurred boundaries, and low contrast remain. To address these issues, we propose CGMU-Net, a classification-guided cross-interaction network. The model integrates a cross-interaction attention fusion module to enhance semantic--boundary feature interaction, a channel-wise feature enhancement module to improve multi-scale consistency, and a classification-guided lesion-aware module to introduce discriminative cues. Experimental results on three public datasets (ISIC 2017, ISIC 2018, and PH2) demonstrate that CGMU-Net achieves Dice scores of 89.21%, 89.91%, and 95.01%, respectively. These results indicate that the proposed method provides accurate and robust lesion segmentation.
Authors
- Bingan Yuan
- ZhengFeng Li (ORCID: https://orcid.org/0009-0006-0769-1483)
- Yijie Tang (ORCID: https://orcid.org/0009-0008-6120-4484)
- BingYan Liu
- MingEn Zhong
Institutions
- Xiamen University of Technology (CN)
- Tsinghua University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-9940905/v1
- Primary Topic
- Cutaneous Melanoma Detection and Management
- Type
- preprint