Enhancing multi-class skin lesion diagnosis through ensemble learning of CNN and transformer architectures

Abstract

Skin cancer remains one of the most prevalent forms of cancer worldwide, highlighting the critical need for accurate and automated diagnostic systems to support early detection and improve patient outcomes. This study presents a deep learning-based framework for multi-class classification of dermoscopic skin lesion images using the HAM10000 dataset. A range of state-of-the-art convolutional neural networks (CNNs) — including DenseNet201, InceptionResNetV2, and Xception — were evaluated under both frozen and fully fine-tuned configurations. Additionally, the performance of Vision Transformer (ViT) architectures was assessed to examine their potential in skin lesion analysis. To enhance classification performance, ensemble learning strategies — namely hard voting, soft voting, and weighted soft voting — were implemented. Experimental results indicate that fully fine-tuned models outperform their frozen counterparts, with InceptionResNetV2(full) achieving the best individual performance with accuracy of 0.88% and F1-score of 0.77%. The highest overall performance was obtained using the proposed weighted soft voting ensemble, yielding an accuracy of 0.89% and an F1-score of 0.80%. These findings demonstrate the effectiveness of ensemble methods and transfer learning based models in advancing automated skin lesion classification. Moreover, the results highlight the potential and limitations of each architecture in clinical applications and provide valuable insights for future research in computer-aided dermatological diagnosis.