A U-Net Segmentation Model for Skin Cancer Based on Multi-scale Channel Fusion Attention

The accurate and efficient segmentation of skin cancer lesions is crucial to the early detection and diagnosis of the condition. However, the often blurred edges and complex color variations in lesion areas complicate the segmentation of the lesions. To address these issues, a U-Net segmentation network based on the multi-scale channel fusion attention (CFA) mechanism (MCFA-UNet) is proposed. The proposed network integrates a multi-scale attention fusion module that independently extracts different scales of local and global features. Furthermore, a multihead channel cross-fusion transformer skip connection module and a CFA module are designed to effectively integrate feature information between skip connections and encoders. In addition, a global feature-extraction module is developed to capture the global information of the input image. Thereafter, MCFA-UNet is validated using two publicly available datasets, ISIC2017 and ISIC2018, and the segmentation metrics, mIoU, DSC, and ACC, reach 84.8%, 91.64% and 96.83%, respectively, demonstrating the excellent segmentation performance of the model.