BGMA-Net: A Boundary-Guided and Multi-attention Network for Skin Lesion Segmentation

In clinical practice, the accurate diagnosis of skin lesions based on medical image segmentation is significant. However, skin lesion image boundaries are often coarse and blurred, and most traditional CNN-based segmentation networks cannot effectively use edge information of shallow layer to guide boundary segmentation of features. To address this problem, we propose a novel neural network called BGMA-Net. This network focuses on the complementarity between edge and object information by combining boundary guidance with multiple attention modules. It integrates boundary and object information, enhances feature representations, and improves skin lesion segmentation accuracy. Specifically, we first propose a simple yet effective boundary-guided attention gate (BGAG) module that integrates local edge information and global positional information to obtain rich boundary information. We further design an effective boundary segmentation attention (BSA) module to refine the boundary information of features. Finally, we propose a channel gated attention fusion (CGAF) module to combine encoder and decoder features, reducing semantic gaps and restoring fine-grained details of target objects. The evaluation of the ISIC 2017 and ISIC 2018 datasets demonstrates that BGMA-Net outperforms state-of-the-art methods, proving the reliability of this framework.