Diagnostic accuracy of artificial intelligence for caries detection on intraoral radiographs: a systematic review and meta-analysis

Background. Deep learning-based artificial intelligence (AI) is increasingly studied for automated caries detection on intraoral radiographs, yet the reproducibility and comparability of published accuracy estimates remain under debate. A rigorous quantitative synthesis strictly focused on intraoral images is needed. Objective — to perform a systematic review and meta-analysis of the sensitivity and specificity of AI algorithms designed to detect carious lesions on intraoral (bitewing and periapical) radiographs. Materia and methods. the review followed the PRISMA 2020 statement. PubMed and eLibrary were searched from January 2020 to April 2025. Two reviewers independently selected prospective/retrospective validation studies and randomized controlled trials (RCTs) evaluating the diagnostic accuracy of AI for caries detection on intraoral radiographs. Methodological quality was assessed with QUADAS-2. A bivariate random-effects model was employed for quantitative synthesis. Results. ten publications were included in the systematic review; five provided data suitable for meta-analysis, and five were analyzed qualitatively. The meta-analysis of five standalone AI validation studies yielded a pooled sensitivity of 0.74 (95% CI 0.71—0.77) and specificity of 0.93 (95% CI 0.89—0.96). Heterogeneity of specificity was high (I²=86%) and largely attributable to Cantu et al. (2020); after its exclusion, specificity increased to 0.96 (95% CI 0.94—0.98) with I²=0%. In the only RCT employing AI assistance (Mertens et al., 2022), sensitivity was 0.81 (95% CI 0.74—0.87) and specificity 0.97 (95% CI 0.95—0.98). Qualitative synthesis of three additional RCTs confirmed that AI support improves clinicians’ sensitivity by 11—31 percentage points, though it may reduce specificity. Conclusions. Standalone AI algorithms demonstrate clinically acceptable accuracy for caries screening on intraoral radiographs, with moderate sensitivity and high specificity. AI-assisted diagnosis is associated with further sensitivity gains, yet requires clinician oversight to minimize overdiagnosis. Multicenter prospective studies with external validation are warranted.