Diagnostic Accuracy of Artificial Intelligence for Dental Caries Detection Across Clinical Imaging Modalities: A Systematic Review and Descriptive Synthesis

Abstract

Background/Objectives: Published estimates of artificial-intelligence (AI) accuracy for dental caries differ in imaging modality, lesion threshold, observational unit, reference standard, and validation design. We mapped the clinical composition and credibility of this evidence and summarized clinician-plus-AI studies separately. Methods: Five databases were searched through 10 June 2026. Eligible diagnostic-accuracy studies used clinically acquired human dental data. QUADAS-3 and a structured GRADE-DTA assessment were applied. Because the studies did not address a common clinical question, no pooled operating point was estimated. Results: A total of 29 reports representing 28 studies were included. Overall risk of bias was high for all 28 estimates; only five studies used external validation. Twelve standalone-AI reports supplied exact, coherent 2 × 2 data. Their sensitivity ranged from 0.360 to 0.940 and specificity from 0.700 to 0.983. The exact-data subset was an availability sample, and most nominal uncertainty estimates could not account for clustering. Two controlled reader studies reported higher sensitivity with AI assistance, accompanied by lower specificity in one study and more invasive treatment decisions in the other. Certainty was very low for both standalone accuracy and incremental clinician benefit. Conclusions: The evidence does not support a transferable accuracy benchmark or a conclusion of net clinical benefit. Locked-model external validation and paired prospective evaluations with patient-relevant outcomes are needed.