Exploring DeiT Transformers for Dermoscopic Image Classification: A Pilot Study

Skin cancer is a significant public health concern, with millions of new cases diagnosed annually worldwide. Early and accurate diagnosis is crucial for effective treatment and improved patient outcomes. Convolutional neural networks have shown promising results in skin cancer classification, but their performance can be limited by the availability of large, high-quality datasets. In this work, we present a novel approach for skin cancer diagnosis using the Data-efficient Image Transformer (DIET) architecture, which has demonstrated state-of-the-art performance on various image classification tasks.. The DeiT architecture is distinguished by its ability to capture long-range dependencies and contextual information, rendering it particularly adept at addressing complex visual patterns prevalent in dermatology. Furthermore, the DeiT model is optimized for performance on small datasets and with limited computational resources, owing to its efficient training methodology and knowledge distillation framework. We evaluate the DeiT model on the ISIC2019 dataset, which encompasses various skin lesion types, including melanoma. Our findings reveal an overall accuracy of 85%, accompanied by high precision and recall rates. These results underscore the potential of the DeiT transformer in advancing diagnostic capabilities in skin cancer diagnosis.