Artificial intelligence analysis of oral mucosal lesions in white‐light clinical photographs: A systematic review of visual features and validation

Artificial intelligence (AI) is increasingly applied to conventional oral photographs for recognition, classification, localization, segmentation, and risk assessment of oral mucosal lesions. This systematic review evaluated how clinically relevant visual information—including color, texture, morphology, borders, and spatial patterns—is represented by AI systems and assessed performance, validation, explainability, human comparison, and clinical translation. PubMed, Embase, Scopus, and Google Scholar were searched for studies directly analyzing conventional white‐light clinical photographs. Risk of bias was assessed with PROBAST + AI, and certainty with Grading of Recommendations Assessment, Development, and Evaluation‐based approaches. Fifty‐three studies were included, 30 published in 2025–2026. Forty‐six used implicit deep‐learning visual representations, three multimodal vision‐language models, two explicit color analyses, and two explicit texture analyses. Thirty‐one studies primarily addressed image‐level diagnostic classification/recognition, 15 spatial detection/localization or lesion delineation, six screening/triage/referral or biopsy‐oriented decision support, and one longitudinal malignant‐transformation prediction. Six used independent external cohorts and five compared AI with human readers. Methodological concerns were frequent, no sufficiently coherent group supported meta‐analysis, and certainty remained limited across evidence bodies. Current evidence therefore supports technical feasibility rather than established clinical effectiveness, with translation requiring stronger patient‐level independence, external validation, prospective multicenter evaluation, and evidence of clinical benefit.