Research on Multi-scale Skin Lesion Segmentation Algorithm Based on Convolution and Transformer
Automatic segmentation of skin lesions is of great significance to assist doctors in clinical diagnosis,treatment,and postoperative observation.Existing convolutions are good at establishing local correlations but cannot capture pixel long-range dependencies,while Tansformer can establish global dependencies of feature information but can cause the loss of local details.Therefore,a multi-scale automatic segmentation network that combines convolution and Transformer is proposed.Firstly,ResNet34 is used as the basic coding block,using its pyramid structure to establish multi-level local correlation of lesions;secondly,the Swin Transformer module is used to capture the long-term dependence of context features.Considering the variable shape and size of lesions,a multi-scale feature aggregation module is proposed to further extract multi-scale information of context features;finally,a decoding block with attention mechanism is used to gradually fuse the multi-level semantic information extracted from the coding block.The experimental results show that the Dice coefficients obtained by the proposed model tested on the ISIC 2017 dataset are as high as 89.55％,and the FPS is as high as 83.Compared with other advanced models,this model has fewer parameters,faster reasoning speed,and higher accuracy.
万方数据
维普
