Lightweight UNet with multi-module synergy and dual-domain attention for precise skin lesion segmentation.
Authors
Affiliations (6)
Affiliations (6)
- School of Management, Xihua University, Chengdu, Sichuan, 610039, China.
- School of Computer and Software Engineering, Xihua University, Chengdu, Sichuan, 610039, China.
- School of Computer and Software Engineering, Xihua University, Chengdu, Sichuan, 610039, China. [email protected].
- Department of Cancer Center, The Second People's Hospital of Neijiang, Neijiang City, Sichuan Province, 641000, China.
- Department of Gastroenterology, People's Hospital of Ningxia Hui Autonomous Region, Ningxia Medical University, Yinchuan, 750002, China.
- Intelligent Policing Key Laboratory of Sichuan Province, Sichuan Police College, Luzhou, Sichuan, 646000, China. [email protected].
Abstract
Skin cancer poses a significant threat to life, necessitating early detection. Skin lesion segmentation, a critical step in diagnosis, remains challenging due to variations in lesion size and edge blurring. Despite recent advancements in computational efficiency, edge detection accuracy remains a bottleneck. In this paper, we propose a lightweight UNet with multi-module synergy and dual-domain attention for precise skin lesion segmentation to address these issues. Our model combines the Swin Transformer (Swin-T) block, Multi-Axis External Weighting (MEWB), Group multi-axis Hadamard Product Attention (GHPA), and Group Aggregation Bridge (GAB) within a lightweight framework. Swin-T reduces complexity through parallel processing, MEWB incorporates frequency domain information for comprehensive feature capture, GHPA extracts pathological information from diverse perspectives, and GAB enhances multi-scale information extraction. On the ISIC2017 and ISIC2018 datasets, our model achieves mIoU and DSC scores of 81.22% and 89.64%, and 81.65% and 89.90%, respectively. These results demonstrate improved segmentation accuracy with low parameter count and computational cost, aiding physicians in diagnosis and treatment. https://github.com/SolitudeWolf/ESMDL-UNet.git.