Abstract

Accurate and interpretable skin lesion classification is essential for early melanoma diagnosis and clinical decision support. This study proposes an explainable computer-aided diagnosis framework based on a Vision Transformer (ViT) for binary classification of benign and malignant skin lesions. To improve model transparency, Grad-CAM and transformer attention maps are integrated through a Dual-Map Fusion strategy, providing complementary local and global visual explanations of the model's predictions. The proposed framework was evaluated using a combined HAM10000 and ISIC dermoscopic image dataset. Experimental results demonstrated an overall classification accuracy of 92.55% and a malignant lesion recall of 94.68%, indicating reliable diagnostic performance with a reduced risk of missed malignant cases. In addition, the fused explanation maps provided more informative and interpretable visual evidence than Grad-CAM or attention maps alone, facilitating a better understanding of the model's decision-making process. The proposed framework combines the strong classification capability of Vision Transformer with enhanced explainability, providing an effective and transparent decision-support tool for automated skin lesion diagnosis. These findings demonstrate the potential of explainable transformer-based models for reliable clinical application in dermatological image analysis.

References

• D. Nicoara, I.C. Pop, M.V. Muntean, R.A. Ilies, R. Nan, P.A. Achimas-Cadariu, The Role of Prevention and Early Detection in Skin Tumors: Correlation Between Educational Level and Tumor Stage at Diagnosis, J. Clin. Med. 14 (2025) 7321. https://doi.org/10.3390/jcm14207321.
• M. Boostani, X. Wortsman, G. Pellacani, K. Füzesi, M. Suppa, V. Del Marmol, F.V. Morandini, J. Perez-Anker, P. Giavedoni, C. Cantisani, L. Boussingault, M. Gyöngy, G. Paragh, K. Avanaki, N. Kiss, Dermoscopy-guided high-frequency ultrasound: Principles and applications in dermatology, JID Innov. 6 (2026) 100446. https://doi.org/10.1016/j.xjidi.2025.100446.
• J. Yoon, J. Han, J. Ko, S. Choi, J.I. Park, J.S. Hwang, J.M. Han, D.D.-J. Hwang, Developing and Evaluating an AI-Based Computer-Aided Diagnosis System for Retinal Disease: Diagnostic Study for Central Serous Chorioretinopathy, J. Med. Internet Res. 25 (2023) e48142. https://doi.org/10.2196/48142.
• I.D. Mienye, T.G. Swart, G. Obaido, M. Jordan, P. Ilono, Deep Convolutional Neural Networks in Medical Image Analysis: A Review, Information. 16 (2025) 195. https://doi.org/10.3390/info16030195.
• I. Afifi, M. Elgendy, M. Abdelfatah, S. El-Sappagh, Vision and convolutional transformers for Alzheimer’s disease diagnosis: a systematic review of architectures, multimodal fusion and critical gaps, Brain Informatics. 13 (2026) 1. https://doi.org/10.1186/s40708-025-00286-7.
• Correction to “Advances in the noninvasive diagnosis of melanoma—40 years beyond the ABCDs,” CA. Cancer J. Clin. 76 (2026). https://doi.org/10.3322/caac.70074.
• M. Ennab, H. Mcheick, Enhancing interpretability and accuracy of AI models in healthcare: a comprehensive review on challenges and future directions, Front. Robot. AI. 11 (2024). https://doi.org/10.3389/frobt.2024.1444763.
• H. Ayaz, E. Sümer-Arpak, E. Ozturk-Isik, T.C. Booth, D. Tormey, I. McLoughlin, S. Unnikrishnan, Post-hoc eXplainable AI methods for analyzing medical images of gliomas (— A review for clinical applications), Comput. Biol. Med. 196 (2025) 110649. https://doi.org/10.1016/j.compbiomed.2025.110649.
• S.A. Almutairi, A dual-stream deep learning framework for skin cancer classification using histopathological-inherited and vision-based feature extraction, Sci. Rep. 15 (2025) 32301. https://doi.org/10.1038/s41598-025-01319-1.