Intelligent Diagnosis and Classification of Skin Disease Based on Medical Images

Skin diseases significantly impact the health of millions of individuals globally. Traditional diagnostic methods, which heavily rely on expert knowledge, suffer from limitations such as being time-consuming and invasive. This paper proposes a deep learning model called D2SM integrating convolutional neural networks (CNN, specifically DSFEM) and Transformer architectures (TCFEM) to achieve efficient and accurate classification of skin diseases. The DSFEM module effectively extracts local spatial features, while the TCFEM module captures global contextual information, jointly enhancing the accuracy and generalization of the classification model. Experimental validation was conducted using the ISIC2020 dataset, and results demonstrate that our approach significantly outperforms existing mainstream models concerning key metrics such as accuracy (82.13 %), recall (87.61 %), and F1-score (82.50 %). Furthermore, interpretability analysis via Grad-CAM confirmed that the model's decision-making focuses appropriately on critical lesion regions, highlighting its strong interpretability and potential for clinical applications.