Skin Lesion Classification Using Vision Transformers on the ISIC 2018 Dataset

One of the most crucial elements in any attempt at increasing the survival of patients is the identification of skin cancer in a timely and accurate fashion. Recent advances in automated medical image analysis have been being pushed by deep learning-based approaches. The present work suggests a multi-class classification of skin lesions with the use of ISIC 2018 and introduces a system built on the Vision Transformer. The proposed model employs the transfer learning concept by finetuning a ViT-Base model on dermoscopic images. An experiment run on an NVIDIA DGX platform in PyTorch and HuggingFace Transformers is carried out. The discriminative capability has been found to be strong in seven categories of skin lesions and has an accuracy of validation of 80.88%. The confusion matrix analysis has also provided a deeper insight into the performance calculation of the classes-wise accuracy, recall, and F1-score. The findings also prove that transformer-based architecture is a promising alternative to classical convolutional neural networks, and its important efficacy in the activity of biomedical image classification.