Abstract
Artificial intelligence has garnered much interest and attention in the last ten years in the field of dermatology. Dermatologists have shown great interest in the ability of deep learning algorithms (especially convolutional neural networks) to accurately classify skin lesions from digital images, frequently approaching or surpassing performance of board-certified dermatologists in laboratory simulations. This critical review of peer-reviewed literature published mainly since 2019 reviewed whether artificial intelligence should be used to replace dermatologists or whether the nuanced integration of humans and artificial intelligence is more suitable. The review explores technical aspects, benchmarking against human observers for skin cancer diagnosis, utility in inflammatory and infectious skin diseases, clinical integration studies, and human perspectives. The results show that artificial intelligence performs well in tasks related to pattern matching of common lesion types in optimal image acquisition settings but falls short of adapting to real-world variability (population, lesion subtypes, benign lesions). Prospective validation assessments show significant degradation when algorithms are applied to practice using different camera models or the photo-types not represented in the training algorithms. Critical appraisal identifies ongoing debates around regulatory, commercial, medico-legal and ethical issues, and the lack of rigorous randomized controlled trials with patient-relevant outcomes. The studies strongly indicate artificial intelligence has potential to serve as a powerful decision support system, to assist and increase the productivity of dermatologists, and to bring dermatologist expertise to remote communities. But the technology is not yet ready to practice independently due to the absence of clinical judgement, integration and learning capabilities. This conclusion builds on the argument that artificial intelligence will not take over the dermatologist's role but dermatologists who harness the power of artificial intelligence may become more successful than those who don't. This review offers guidance for clinicians, researchers and policy makers to come together to implement the ethical use of artificial intelligence in skin disease diagnosis.