Skin Lesion Classification Using Machine Learning and Deep Learning: A Narrative Review of Datasets, Methods, Clinical Evidence, and Open Challenges

Skin cancer is among the most prevalent cancers in the world, and early detection of melanoma is the best strategy to lower the associated mortality. Computer-aided diagnosis of skin lesions evolved from hand-crafted feature pipelines based on the ABCD rule and texture descriptors into deep neural networks achieving dermatologists' performance on benchmark-controlled tasks. In this review, we provide a full overview of the automated skin lesion classification pipeline: from the clinical background, over public skin lesion image datasets and challenges, and the early approach of hand-crafted computer vision descriptors over modern state-of-the-art deep learning architectures, vision transformers, self-supervised, and foundation models, multimodal skin lesion analysis explainability uncertainty estimation, down to reader studies or other comparative analyses between algorithms and clinicians. We provide a critical reflection on how far benchmark performance is from translational readiness, and address problems of dataset bias and under-representation of skin types; spurious correlations and artefacts; class imbalance; domain shift; reporting quality; privacy; and regulation. Finally, we conclude with a shared vision for future research in the field of skin cancer detection including the curation of human-labelled representative datasets; use of prospective validation studies; human-AI partnership systems; private training data; and responsible, explainable models. This is a narrative review article and not a systematic review and meta-analysis. We have described results of articles at the study level, and note that differences across performance metrics should be put into context of differences across datasets, reference tasks, and validation approaches.