Research on Dermatological Classification Algorithm Based on the Fusion Model*

A new dermatology classification algorithm is proposed to use CNN instead of Patch Embedding part of Swin Transformer for local feature extraction. The method gives full play to the advantages of CNN in extracting local features, while exploiting the excellent ability of Transformer in extracting global features, realizing the complementary advantages between the two. In addition, the SimAM self-attentive mechanism is introduced in the CNN part, and after training on the ISIC2019 dataset, the fused network achieves an accuracy of 90.6%, which improves the performance by 7.4% and 2.1% relative to the methods applying ResNet and Swin Transformer independently. These results demonstrate the significant advantages of the fused algorithm in the field of dermatology classification.