Deep Learning and Explainable AI for Skin Cancer Classification: A Comparative Review of CNN, Transformer, and Hybrid Architectures

Abstract

Skin cancer continues to account for numerous cancer mortality cases around the globe, making the timely and accurate diagnosis essential for achieving optimal outcomes from its treatment. Deep learning research has made significant strides in this field, leading to the development of various architectures, including Convolutional Neural Networks (CNNs), Vision Transformers (ViT), hybrid CNNs and transformers, and explainable methods that can assist in the automated skin cancer classification. In this review paper, a detailed comparative analysis of these approaches will be conducted based on criteria such as the architecture of the model, datasets used, classification accuracy, computational requirements, advantages, drawbacks, and explainability methods. Almost all reviews in the area analyze either CNNs or Transformers or explainable methods individually, making it difficult to assess how they compare in practical application. For this reason, this review provides a unified assessment of all the aforementioned approaches and focuses on their areas of practical applicability in the context of skin cancer classification. Additionally, possible future directions of development and necessary changes for making these solutions clinically applicable will be discussed.