A Multi-Scale Boundary-Aware Attention Network for Skin Cancer Classification

作者：Jianxiang Deng, Ruofei Wang, Yufu Wu, Jiarui Chen, Jin Lin · 年份：2025 · DOI：10.1109/iccvit67848.2025.11391446 · 研究领域：Cutaneous Melanoma Detection and Management、Nonmelanoma Skin Cancer Studies、Face recognition and analysis

Among all malignancies, skin cancer ranks as highly prevalent and lethal, making timely and precise diagnosis essential for patient well-being. Conventional manual screening methods, however, suffer from low efficiency. Artificial intelligence has gained increasing adoption in dermatological lesion detection over recent years, though current deep learning approaches still require accuracy improvements. Addressing these limitations, this work introduces the Multi-scale Boundary-aware Semantic Attention (MBS-Attention) mechanism, a novel architecture that achieves superior trade-offs between classification performance and computational cost for deep learningdriven skin cancer recognition. The proposed architecture combines three components: a semantic-aware component, a multi-scale feature enhancement component, and a boundary-aware component, collectively addressing challenges posed by varying lesion scales and unclear lesion boundaries. Extensive experiments conducted on the ISIC2017 dataset demonstrate that when MBS-Attention is incorporated into the efficient ConvNeXt-Tiny architecture, the classification accuracy reaches$\mathbf{8 1. 1 7 \%}$with an F1-score of 74.26%. These results show that our method outperforms models of similar complexity while incurring a modest overhead$(+1.81 \mathrm{M}$params, +0.74 G FLOPs), making it a practical option for clinical use.