Fusion multi-scale Transformer skin lesion segmentation algorithm

To address the problem of lack of multi-scale feature extraction in existing skin lesion image segmentation，which leads to lack of detailed information and incorrect segmentation of skin lesion regions，this paper proposes a fusion multi-scale Transformer encoder-decoder network skin lesion segmentation algorithm. First，a hierarchical encoder is constructed using Transformer Block，which analyses the skin lesion region from the perspective of global feature variation at multiple scales. Then，the multi-scale fusion module，channel attention module and concat layer are used to construct the fusion decoder. The multi-scale fusion module fuses shallow network information and deep network information in the hierarchical encoder to enhance the dependency between spatial and semantic information，and the channel attention module can effectively identify channels containing rich feature information and improve the segmentation accuracy of the algorithm. Finally，an expansion module is introduced to recover the image size to meet the practical requirements. The proposed algorithm was experimentally tested on three public datasets，ISBI2016，ISBI2017 and ISIC2018. The pixel accuracies were 96.70%，94.50% and 95.39%，respectively，and the mean intersection over union were 91.69%，85.74% and 89.29%，respectively，with the overall performance of the tested algorithms outperforming existing algorithms.Simulation experiments show that the multi-scale Transformer encoder-decoder network can effectively segment skin lesion images，providing a new window for the diagnosis of modern skin diseases. © 2024 Editorial Board of Jilin University. All rights reserved.