Diagnostic Performance of Artificial Intelligence for the Detection and Segmentation of Periapical Lesions on Cone-Beam Computed Tomography: A Systematic Review and Meta-Analysis

Background/Objectives: Cone-beam computed tomography (CBCT) improves periapical lesion detection, but its interpretation is expertise-dependent. This systematic review evaluated artificial intelligence (AI) for detecting, classifying, and segmenting periapical lesions on CBCT. Methods: PubMed, the Cochrane Library, Scopus, Web of Science, and Google Scholar were searched to 12 December 2025 (PROSPERO: CRD420251184440). Studies comparing AI with an expert- or annotation-based reference standard on human CBCT scans were eligible. Risk of bias was assessed with the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool, and certainty of evidence with the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach. Sensitivity and specificity were pooled with a bivariate model where complete 2 × 2 data were available, and all metrics with univariate random-effects models. Results: Eighteen studies were included (detection or classification, 16; segmentation, 6); 15 had unclear risk of bias in at least one domain, and most were retrospective and internally validated. In the bivariate model (seven studies), sensitivity was 88.2% (95% confidence interval [CI]: 77.6–94.2%) and specificity 85.1% (73.4–92.1%). Univariate pooled sensitivity was 90.9% (84.2–94.9%; 14 studies), specificity 89.7% (83.2–93.9%; 11 studies), accuracy 86.7% (78.9–92.0%; 10 studies), and F1-score 83.7% (76.4–89.1%; seven studies). The pooled lesion Dice similarity coefficient was 0.70 (0.61–0.78; six studies), indicating moderate spatial overlap with reference annotations. Heterogeneity was considerable for detection metrics (I2 = 90–96%), prediction intervals were wide (sensitivity 48–99%), and the certainty of evidence was very low. Conclusions: AI shows promising performance for CBCT-based periapical lesion detection under research conditions, whereas segmentation overlap is moderate. Equivalence to clinicians and clinical effectiveness are not established; prospective, externally validated studies with standardized reference standards are needed.