Skin lesion segmentation by fusing local and global features using axial shift and spatial state model

Skin lesions present challenges to medical diagnosis due to their complex features, including shape variations, uneven color, and blurred boundaries. Currently, models based on convolutional neural networks (CNNs) and Transformers often have too many parameters, making them difficult to deploy in resource-limited medical environments while also struggling to balance local and global features. To address this, this paper proposes a Shift-Mamba structure that effectively captures local features through an axial shift mechanism and fuses global features using Mamba's spatial state model (SSM). Notably, the new model (SM-UNet) designed based on the Shift-Mamba structure has only 0.02 million (M) parameters, making it one of the lightest models available, much lighter than those based on CNN or Transformer architectures. The SM-UNet model was validated on the ISIC 2017 and ISIC 2018 datasets, achieving IoU and Dice scores of 84.04%, 91.15% and 82.50%, 90.23%, respectively. These results surpass those of existing segmentation models, demonstrating the superiority of SM-UNet in the task of skin lesion segmentation. Code is available at https://github.com/guangguangLi/SM-UNet.