RsCU-Net: Multi-scale Features Based U-Net for Skin Lesion Segmentation

Aiming at the diversity, blurred edges and hair noise in skin lesion areas, a multi-scale feature network (RsCU-Net) with U-Net as the baseline model was proposed. Firstly, the CA module is introduced to enhance the network's spatial awareness. Secondly, the RS module is incorporated to extract multi-scale features and adaptively adjust the weights of each feature channel. Finally, the fusion of BCE loss function and Dice loss function addresses the issues of class imbalance in skin lesion image samples and enhances difficult sample mining. We validated our proposed network on the ISIC2018 dataset. Compared with the baseline U-Net, RsCU-Net demonstrated respective improvements of 3.42%, 2.14%, and 4.92% in MIoU, Dice score, and Precision metrics. Additionally, it achieved reductions of 0.175 in HD95 and 0.1126 in RVD.