Design and Development of a Vision Transformer (ViT) Model for Predicting Skin Cancer from Skin Images

This study aims to develop an AI model using Vision Transformer (ViT) for more accurate melanoma detection. The ViT model is compared to a traditional Convolutional Neural Network (CNN) model to evaluate improvements in detecting skin cancer. Materials and Methods: The study had involved two different models: Group 1: Vision Transformer model feature capturing high-resolution images of melanoma taken from ISIC set. Group 2: conventional CNN model having been trained on the same ISIC set. The images have been resized, normalized, and preprocessed while both models are loaded with a learning strategy that adapts learning rate. Performance evaluation was done using the accuracy estimation of senescence and absorbance along with calculating area under the curve (AUC) for 10 rounds. Result: The ViT model outperformed the CNN model, with accuracy ranging from 94.15% to 95.17 %, while the CNN model's accuracy ranged from 70 % to 90 %. The ViT model also showed more consistency, with a much lower variation in results. The performance difference was statistically significant (p < 0.05). Important features for detection included the shape and texture of lesions. Conclusion: The Vision Transformer model yielded better accuracy than the CNN model in melanoma detection and presented consistent results. This proposes another idea for employing Vision Transformer in telemedicine for the early detection and individualized treatment. Further studies will have to include model interpretability for practice in the real world.