An Enhanced Skin Cancer Detection using Medical Transformers

Skin cancer is among the most common and hazardous types of cancers, and the issue of early and accurate diagnosis is inherent in the context of raising the survival rates. Improved models are proposed in this paper in order to classify skin cancer on dermoscopic skin using a better Vision Transformer (ViT). The proposed ViT model consists of patch-wise tokenization, multi-head self-attention, and positional embedding (in contrast to the existing convolutional neural networks (CNNs) that are limited by the capacity to reveal the long-range dependencies) to learn the informative global images of medical images. Two benchmark datasets, i.e. ISIC-2019 Balanced Subset and ISIC9042 Full Dataset are used to train and test the model with a classification accuracy of 88.2 and 88.5 percent, respectively. Grad-CAM is integrated to be capable of visualizing where the model concentrates to as a way of interpreting it and relying on its clinical applications to establish the regions of the injury that lead to the classification. The applicability of transformer based architecture to real world dermatological diagnostic can be explained by the fact that the architecture has been proven to achieve positive performance improvement in the accuracy of the diagnostic and the stability of the system relative to the latest state of art approaches and methods.