Early skin cancer detection can significantly increase patient survival. Deep learning models are frequently employed in computer-aided diagnosis (CAD) to assist in the classification, segmentation, and recognition of skin lesions. Despite their effectiveness, convolutional neural networks (CNNs) may have trouble comprehending the significant variations in lesion shapes (intra-class variation) and the similarities between various lesion kinds (inter-class similarity). Furthermore, significant local information in dermoscopic pictures may be lost due to CNNs’ extensive downsampling. Transformer-based models have recently demonstrated impressive results in medical image analysis by using self-attention to capture both local and global data. In this study, we use the HAM10000 dataset to provide a Vision Transformer model for multi-class skin lesion categorization. Our method uses ImageNet transfer learning to enhance performance on sparse medical data. The model shows promise for accurate skin lesion categorization with a validation accuracy 91.76%. For AI applications pertaining to dermatology, this paper shows how transformer structures can be useful alternatives for conventional CNNs.