Artificial Intelligence in the Diagnosis of Oral Diseases: Current Applications, Diagnostic Accuracy, and Future Perspectives

Abstract

Background: Artificial intelligence (AI), particularly machine learning and deep learning, is increasingly investigated as a decision-support technology for oral diagnosis. Dentistry is well suited to AI because diagnosis routinely integrates radiographs, photographs, three-dimensional imaging, pathology, and structured clinical information. Objective: This evidence-based literature review synthesizes current applications of AI in oral diagnosis, examines reported diagnostic performance, identifies methodological limitations, and outlines requirements for safe clinical translation. Methods: A structured narrative review was conducted using peer-reviewed literature identified through targeted searches of PubMed-indexed records and publisher databases, emphasizing systematic reviews, meta-analyses, and contemporary reporting and governance guidance. Reported performance measures were extracted without generating new patient-level estimates. Results: Evidence is most mature for dental caries, oral potentially malignant disorders and oral cancer, periapical lesions, periodontal disease, and automated tooth identification. Recent syntheses report promising performance, including accuracy ranges of 71–99% for caries models, 85–100% for deep-learning oral-cancer studies, 74–100% for oral mucosal lesion detection, pooled sensitivity/specificity of 0.94/0.96 for periapical radiolucencies, and pooled dental-diagnostic sensitivity/specificity of approximately 0.85/0.93 across low-bias systematic reviews. However, heterogeneity, limited external validation, single-center datasets, spectrum bias, inconsistent reference standards, and uncertain clinical benefit remain major barriers. Conclusion: AI can augment oral diagnosis, but current evidence supports clinician-supervised decision support rather than autonomous diagnosis. Multicenter prospective validation, transparent reporting, representative datasets, explainability, equity safeguards, and workflow-level outcome evaluation are essential before widespread implementation.