A vision transformer-based domain-adapted framework for skin cancer detection

Skin cancer is a major dermatological health concern in which early and accurate detection can support timely clinical evaluation and treatment. Artificial intelligence, particularly deep learning, has increasingly been investigated for automated skin-lesion classification using dermoscopic and clinical images. Convolutional Neural Networks (CNNs) have traditionally provided strong image representations, while Vision Transformers (ViTs) have introduced self-attention-based mechanisms capable of modeling relationships between image regions. However, performance reported on an individual dataset does not necessarily translate to reliable performance on an external dataset. Differences in imaging devices, illumination, image resolution, acquisition protocols, patient populations, lesion characteristics, class distributions, and imaging modalities can introduce domain shift and reduce cross-dataset generalization. This review examines the evolution of deep-learning approaches for skin-lesion classification, with emphasis on CNNs, transfer learning, attention mechanisms, Vision Transformers, lightweight Transformers, hybrid CNN–Transformer models, multimodal learning, federated learning, and domain-adaptation strategies. Particular attention is given to cross-dataset evaluation and approaches such as domain-adversarial learning, unsupervised domain adaptation, selective alignment, test-time adaptation, and conformal ensemble methods. The review identifies persistent research gaps related to dataset dependency, computational cost, limited target-domain labels, inconsistent evaluation protocols, class imbalance, interpretability, and deployment robustness. Based on these gaps, a future research direction involving a lightweight Vision Transformer with domain adaptation for cross-dataset skin-lesion classification is proposed. The proposed direction is presented as future experimental work, and no experimental results are claimed in this review.