Abstract:In skin lesion segmentation, challenges such as blurred boundaries, irregular lesion morphologies, and multi-scale variations are commonly encountered. To address these issues, this study proposes a skin lesion segmentation method based on adaptive attention and multi-scale feature fusion. The proposed framework is mainly composed of an adaptive synergistic attention (ASA) module and an enhanced group aggregation bridge (EGAB) module. The ASA module performs refined calibration of multi-level features through a synergistic channel-spatial attention mechanism, frequency-domain global context modeling, and explicit boundary enhancement. The EGAB module improves feature fusion quality by incorporating cross-scale attention mechanisms and grouped multi-scale dilated convolutions. Experimental results on the ISIC-2017 and ISIC-2018 datasets demonstrate that the proposed method achieves superior segmentation accuracy compared with several mainstream approaches, while maintaining model efficiency and significantly enhancing segmentation performance for lesions with blurred boundaries and complex morphologies.