A novel CNN-ViT-based deep learning model for early skin cancer diagnosis

Abstract

Skin cancer is a serious global health issue where early detection is crucial for effective treatment and improved patient outcomes. However, accurate diagnosis is challenging due to the variety of subtypes and imaging complexities. This study introduces an innovative deep learning model based on the MetaFormer architecture, optimized specifically for skin cancer. The Proposed Model features a hybrid design that replaces traditional self-attention methods with novel focal self-attention mechanisms, enhancing its ability to identify critical regions, reduce noise, and extract features more effectively, ultimately boosting diagnostic accuracy. To evaluate the model’s generalization capabilities, it was tested on two benchmark datasets: ISIC 2019, which includes a diverse set of dermatological images across eight skin cancer classes, and HAM10000, widely used in dermatological research. The model achieved outstanding results, including an accuracy of 0.9254, precision of 0.9041, recall of 0.8768, and an F1-score of 0.8886 on ISIC 2019, and an accuracy of 0.9501, precision of 0.9470, recall of 0.9211, and an F1-score of 0.9334 on HAM10000. The Proposed Model surpasses existing methods in the field, outperforming ten advanced CNN models and twenty state-of-the-art ViT models under the same training and evaluation conditions. With a lightweight design of just 35.01 million parameters, it is optimized for real-time and mobile applications, making it highly practical for clinical use. Its reliable performance ensures accurate diagnoses, which are essential for early intervention and treatment, addressing a critical need in modern healthcare.