Comparative Analysis of Single and Hybrid CNN-Transformer Models for Melanoma Detection Using the ISIC Dataset

Early detection of skin cancer, particularly melanoma, is crucial for improving patient outcomes. This study presents a comprehensive comparative analysis of deep learning architectures for automated skin lesion classification. This study evaluated eight models across four architectural categories: traditional convolutional neural networks, residual networks, vision transformers and novel hybrid approaches combining CNNs with transformers. Using the International Skin Imaging Collaboration dataset with 3,267 dermatoscopic images, the study assessed both classification performance and computational efficiency. Hybrid architectures consistently outperformed individual models, with the ResNet+ViT hybrid achieving the highest accuracy of 85.1% and AUC of 0.932. However, these performance improvements came with significant computational costs, highlighting important trade-offs for practical deployment. MobileNetV2 offered the best efficiency while maintaining a reasonable accuracy of 79.1%, making it suitable for resource-constrained environments. The findings demonstrate that combining complementary feature extraction mechanisms from CNNs and transformers creates more robust representations for skin lesion classification, approaching the diagnostic accuracy of dermatologists reported in the literature. This research contributes valuable insights for the development of automated diagnostic support tools that could improve early detection rates and enhance patient outcomes in skin cancer management.