An Improved Vision-Transformer Network for Skin Cancer Classification

The early detection of skin cancer through automation is crucial for enhancing patient recovery prospects. In this study, we present an innovative approach for classifying skin cancer lesions using a Vision transformer (ViT) and evaluate it on the International Skin Imaging Collaboration (ISIC) 2017 dataset. The evolution of computer vision has led to the emergence of ViT, which possesses a unique ability to detect intricate patterns and features through self-attention mechanisms. This allows ViT to recognize extensive dependencies within images, resulting in performance exceeding conventional CNN models. In comparison with the current state-of-the-art Inception-ResNet-V2 + Soft Attention (IRV2 + SA) technique, our proposed model exhibits superiority in accuracy, precision, recall, and AUC-ROC score for binary classification tasks in the ISIC 2017 challenge. Furthermore, the method demonstrates credibility as a reliable tool for lesion classification. The outcomes underscore ViTs’ potential as a promising alternative to established convolutional neural network architectures for skin cancer lesion categorization. https://github.com/Gayathri-Shajimon/Skin-cancer-lesion-classification-using-ViT