MWAG-Net: multi-scale wavelet enhancement and attention gating network for skin lesion segmentation
Physica Scripta · 15.05.2026 · 101 (19) · 195004
DOI: 10.1088/1402-4896/ae6735 ↗Kurz gefasst
Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 13.05.2026. Verfasst von Wenli Fu und 2 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben. Gefördert von: Hunan Provincial Natural Science Foundation of China, ”14th Five-Year Plan” Key Disciplines and Application-oriented Special Disciplines of Hunan Province, Science and Technology Plan Project of Hunan Province.
Abstract (Originalsprache)
Accurate skin lesion segmentation is essential for early melanoma detection and clinical decision-making. However, this task remains challenging due to blurred boundaries, low contrast, and irregular shapes. Existing convolutional neural network methods show limited global modeling capability. Approaches using wavelets or attention mechanisms fail to fully leverage high-frequency information for explicit feature refinement, leading to suboptimal performance on ambiguous boundaries. To address these issues, we propose a novel Multi-scale Wavelet Enhancement and Attention Gating Network (MWAG-Net), which mainly consists of the Wavelet-Enhanced Dual-branch Convolution (WED-Conv) and Wavelet-Guided Attention Gating (WGAG) modules. First, the WED-Conv module performs multi-level discrete wavelet decomposition to separate features into high-frequency and low-frequency components. Then, these components are enhanced through residual augmentation to jointly model blurred boundaries and global structure. In addition, the WGAG module uses wavelet-derived high-frequency details as explicit guidance signals. By integrating spatial and channel attention, WGAG adaptively refines boundary-related encoder features for precise boundary reconstruction. Experiments on the ISIC 2017, ISIC 2018, and PH2 datasets show that MWAG-Net achieves the best overall performance. Paired t-tests verify the statistical significance of its improvements. Compared with UWT-Net, MWAG-Net improves the Jaccard Index (JI) by 2.17% and 1.07%, and reduces the 95th Percentile Hausdorff Distance (HD95) by 1.64 mm and 1.11 mm on ISIC 2017 and ISIC 2018, respectively. The algorithm can provide a reliable solution for the early diagnosis of skin cancer.
Förderung
- Hunan Provincial Natural Science Foundation of China
- ”14th Five-Year Plan” Key Disciplines and Application-oriented Special Disciplines of Hunan Province
- Science and Technology Plan Project of Hunan Province
Zitieren
APA
Fu, W., Deng, H., & Ye, J. (2026). MWAG-Net: multi-scale wavelet enhancement and attention gating network for skin lesion segmentation. Physica Scripta, 101(19), 195004. https://doi.org/10.1088/1402-4896/ae6735
Vancouver
Fu W, Deng H, Ye J. MWAG-Net: multi-scale wavelet enhancement and attention gating network for skin lesion segmentation. Physica Scripta. 2026;101(19):195004. doi:10.1088/1402-4896/ae6735
BibTeX
@article{Fu2026,
author = {Fu, Wenli and Deng, Hongwei and Ye, Jixun},
title = {MWAG-Net: multi-scale wavelet enhancement and attention gating network for skin lesion segmentation},
journal = {Physica Scripta},
year = {2026},
volume = {101},
number = {19},
pages = {195004},
doi = {10.1088/1402-4896/ae6735}
}
EUVIMED