Abstract: In the treatment of skin cancer, early skin lesion segmentation technology is of great importance. However, in existing skin lesion segmentation models, global features of images are difficult to be extracted by CNNs. Computational overhead of the model is increased by Transformers, resulting in persistently high training costs. To solve this problem, the Local Feature Enhanced Mamba U-Net (LEM-UNet) model is proposed in this paper. Based on Vision Mamba-UNet, the decoder is reconstructed to enhance global feature extraction. A multi-scale feature extraction residual module (RMS) and a local feature extraction module (LFM) are proposed. These modules strengthen the model’s ability to segment lesions of different sizes in images and improve the model’s accuracy in capturing and extracting lesion boundary features, thus enhancing the overall segmentation effect of the model. Comparative experiments are conducted on the ISIC2017 and ISIC2018 public skin lesion segmentation datasets. The results show that compared with the original baseline model, mIoU, DSC, Acc, Spe, and Sen increase by 0.97, 0.56, 0.22, 0.20, 0.29 percentage points and 1.55, 0.91, 0.41, 0.16, 1.23 percentage points respectively. The model performs particularly well in scenarios where the contrast between the segmented object and the background is not high.

Key words: skin lesion segmentation, Mamba, local features, global features, multi-scale