Machine learning for clinical decision support in cone-beam CT: a systematic review and meta-analysis of performance in oral and maxillofacial imaging

Cone-beam computed tomography (CBCT) is widely used in oral and maxillofacial practice, yet interpretation can be challenging because of anatomical complexity and modality-specific artifacts. Machine-learning (ML) methods are increasingly applied across diagnostic and technical CBCT tasks; however, reported performance is heterogeneous, and previous quantitative synthesis has covered a smaller and narrower evidence base. This systematic review and meta-analysis followed PRISMA 2020 and incorporated applicable PRISMA-DTA items. The protocol was prospectively registered on the Open Science Framework. PubMed, Web of Science, and Scopus were searched from inception through April 1, 2026. Random-effects meta-analyses were performed separately for accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). The unit of synthesis was the study-level reported performance estimate. For one study lacking a reported AUC uncertainty estimate, variance was approximated using available test-set information; the pooled AUC was therefore interpreted with additional caution. Leave-one-out analysis assessed robustness of the pooled accuracy estimate. Seventy-four studies were included. Eligible subsets contributed to the quantitative analyses: 8 studies for AUC, 30 for accuracy, 21 for sensitivity, and 9 for specificity. The pooled AUC was 0.93 (95% CI, 0.88–0.98; I²=40.7%), accuracy was 0.88 (95% CI, 0.87–0.89; I²=96.1%), sensitivity was 0.95 (95% CI, 0.94–0.96; I²=95.5%), and specificity was 0.88 (95% CI, 0.86–0.89; I²=98.1%). Leave-one-out analysis showed that pooled accuracy was not driven by any single study. Considerable heterogeneity limited direct clinical generalizability. ML models demonstrated generally high reported performance across diverse oral and maxillofacial CBCT applications. However, substantial clinical and methodological heterogeneity, predominantly internal validation, inconsistent evaluation units, and limited assessment of calibration and clinical utility restrict generalizability. Summary estimates should therefore be interpreted as field-level averages rather than expected performance for a single model or clinical application. Multicenter external validation, standardized reporting, calibration assessment, and prospective clinical evaluation remain priorities. Overview of the systematic review and meta-analysis of machine-learning applications in cone-beam computed tomography for oral and maxillofacial imaging. The review included 74 studies covering detection, classification, anatomical segmentation, localization, treatment-planning support, and image-enhancement tasks. Random-effects meta-analysis generated summary estimates for accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve. Although the pooled results indicated generally high performance, considerable heterogeneity was observed for accuracy, sensitivity, and specificity, and the AUC synthesis was considered exploratory because the uncertainty estimate for one contributing study was approximated using available test-set information. The evidence was predominantly derived from retrospective, internally validated, and single-center datasets, with limited assessment of calibration or prospective clinical utility. These findings indicate a potential clinician-supervised decision-support role but do not establish readiness for autonomous or routine clinical implementation.