Improving Skin Lesion Diagnosis with a Hybrid Deep Learning Framework: A Study Using ISIC Dermoscopic Images

Skin lesions are a significant public health concern, underscoring the need for reliable and accurate diagnostic tools for early cancer detection. This paper presents a Deep Learning (DL) framework which utilizes DenseNet201 and Vision Transformer (ViT) architectures to improve skin lesion classification. Using the International Skin Imaging Collaboration (ISIC) skin cancer dataset, we applied data augmentation, regularization, and adversarial training to improve the generalizability and resilience of the model. The ensemble model outperformed DenseNet201 and ViT individually, achieving validation metrics of 99.50% accuracy, 99.55% precision, 99.52% recall, and 99.53% F1-score, along with a Cohen's Kappa coefficient of 0.99. These results highlight the ensemble's potential for reliable, high-accuracy clinical skin cancer diagnostics.