LEVit-Skin: A balanced and interpretable transformer-CNN model for multi-class skin cancer diagnosis

Skin cancer is a major cause of death, making early detection essential. This study presents LEVit, an explainable and class-balanced deep learning framework designed for multiclass skin lesion classification. LEVit combines a hybrid Vision Transformer (ViT) with a Convolutional Neural Network (CNN). We evaluated LEVit on two benchmark dermoscopic datasets: HAM10000, which consists of 10,015 images across 7 classes, and ISIC 2019, with 25,331 images spanning 8 classes. Both datasets have notable class imbalances. To address this issue, we applied advanced augmentation techniques to oversample minority classes, ensuring a uniform class distribution and enhancing the model's ability to generalize. LEVit effectively captures local lesion textures and global spatial relationships through its integrated self-attention and convolutional modules. We compared its performance against four state-of-the-art models: NASNet, SqueezeNet, SE-Net, and Xception, across four metrics: F1 Score, Specificity, Matthews Correlation Coefficient (MCC), and Precision-Recall Area Under the Curve (PR AUC). LEVit achieved outstanding results, with a F1 Score of 98.11% and a PR AUC of 98.57% on the ISIC 2019 dataset, and a F1 Score of 96.11% and a PR AUC of 96.62% on HAM10000. For interpretability, we utilized Grad-CAM to generate class-specific heatmaps, which highlight the key areas of lesions that influence the model's predictions. This work demonstrates that balanced training and a hybrid architecture can enhance both classification accuracy and interpretability in skin cancer diagnostics, effectively addressing the limitations of existing models and paving the way for reliable clinical applications.