Abstract

The increasing number of skin cancers underscores the critical importance of early detection and accurate classification to improve treatment outcomes. Melanoma, a malignant skin cancer, has the highest mortality rate among all skin cancer types. Early detection of melanoma significantly enhances the chances of effective treatment and survival rates. This research presents a comparative analysis of cutting-edge deep learning methodologies in medical imaging, specifically focusing on Vision Transformers (ViT) and Convolutional Neural Networks (CNNs) for melanoma cancer detection. This study further examines the influence of domain-specific transfer learning on improving melanoma detection accuracy by pre-training these deep learning models on various datasets, such as ImageNet, BreakHis, and ISIC 2019. The models are then meticulously fine-tuned using a private annotated dataset of melanoma dermoscopic images. In addition, we employed the k-fold cross-validation technique to evaluate the reliability of our models. Our experimental results highlight the significant performance of advanced deep learning methodologies and transfer learning approaches, with the ViT-B16 model achieving an exceptional diagnostic accuracy of 97.97%, outperforming other models, specifically the pre-trained CNNs models. Moreover, This study highlights the critical role of large, diverse datasets in transfer learning, demonstrating their effectiveness in improving model performance for melanoma detection.