Lesion boundary detection for skin lesion segmentation based on boundary sensing and CNN-transformer fusion networks.

Traditional convolutional neural networks often struggle to capture global information and handle ambiguous boundaries during complex skin lesion segmentation tasks. To tackle this challenge, we proposed MPBA-Net, a hybrid network that integrates multi-pooling fusion and boundary-aware refinement. The network integrated Convolutional Neural Network (CNN) and Transformer to generate rich skin lesion feature maps for comprehensive feature extraction. Specifically, we introduced a boundary-aware attention gate (BAAG) module in the Transformer encoder layer and added a boundary cross attention (BCA) module at the end of the network to capture critical skin lesion boundary features. Additionally, we developed a multi-pooling fusion (MPF) module that extracts global multi-scale features by fusing improved Spatial Pyramid (SP) and Atrous Spatial Pyramid Pooling (ASPP). To optimize training, we designed a Point Loss derived from Binary Cross-Entropy (BCE) and combined it with Dice Loss to form a hybrid loss function. This approach not only enhances classification performance but also provides more precise measurement of the similarity between segmentation results and ground truth annotations. Ablation experiments on the ISIC2018 dataset validated the effectiveness of our fusion strategies and network improvements. Comparative experiments on the ISIC2016, ISIC2017, and ISIC2018 datasets showed that the Dice index of MPBA-Net outperformed other comparative segmentation methods in all three datasets, achieving 91.47 %, 87.04 %, and 88.93 %, respectively. Quantitative and qualitative results demonstrate that our method improves skin lesion segmentation accuracy, aiding dermatologists in clinical diagnosis and treatment. Our code is available at https://github.com/FengYuchenGuang/MPBA-Net.