Performance enhancement of visual transformer models via dual-module collaboration for skin tumor classification

This study develops and evaluates an artificial intelligence-based method for classifying skin tumors using clinical and dermoscopic images across seven benign and malignant types. Building on this dataset, we propose an optimized Vision Transformer (ViT) architecture incorporating two key modules: a data preprocessing module, CrossNorm and SelfNorm (CNSN), and a feature adaptation module, M ulti-c o g n itive Visual A dapter (MONA), to enhance model performance and generalization. Comparative and ablation experiments were conducted, with rigorous evaluation using accuracy, precision, recall, F1-score, and area under the curve (AUC). Results demonstrate that the improved ViT achieves a classification accuracy of 92.14%, surpassing other neural network models by 0.13–18.75 percentage points, with precision, recall, F1-score, and AUC of 88.27%, 84.80%, 86.41%, and 99.38%, respectively. Ablation studies show that integrating CNSN and MONA increases accuracy, precision, recall, F1-score, and AUC by 1.12, 1.63, 3.48, 2.71, and 0.03 percentage points compared with the baseline ViT. The proposed model effectively supports physicians in diagnosis, offering more precise and efficient decision-making and holds significant potential for advancing skin tumor diagnostic technologies.