Optimized Vision Transformer with Adaptive Attention for High-Precision Skin Cancer Classification

This work presents an effective Vision Transformer model with Adaptive Learning and Normalization (ViT-ALN), applied to skin cancer classification with the HAM10000 set. To enhance the feature extraction and learning, the authors employ GELU activation, Layer Normalization, and Dropout in the construction of ViT-ALN. The model was trained on re-scaled and augmented images and the model successfully established a classification protocol. In this study, ViT-ALN has a classification accuracy of 96.8% higher than most traditional CNN architectures. These were performed over 250 iterations and led to marked improvements in the ability to distinguish between benign and malignant skin lesions. From these results, it is corroborated that ViT-ALN architecture improvement brings better diagnostic performance for serving the clinical need in dermatology.