Artificial intelligence for detecting orthodontic root resorption: A systematic review and meta-analysis of diagnostic accuracy.
Authors
Affiliations (2)
Affiliations (2)
- Department of Dentistry, University of São Paulo, São Paulo, Brazil. Electronic address: [email protected].
- Department of Dentistry, Federal University of Minas Gerais, Belo, Horizonte, Brazil.
Abstract
External root resorption (ERR) and external cervical resorption (ECR) are common orthodontic complications with prognostic impact. Early, accurate detection helps prevent irreversible damage. Our objective is to synthesize diagnostic accuracy of artificial intelligence (AI) for ERR/ECR on orthodontic imaging and compare performance by imaging modality and model architecture. Data Sources: Web of Science, PubMed, Scopus, Embase, IEEE Xplore, and Cochrane Library (through September 9, 2025). Original human or ex-vivo imaging studies of AI models for ERR/ECR with extractable accuracy data. Two reviewers independently extracted TP/FP/TN/FN. Pooled sensitivity/specificity and HSROC were estimated via a bivariate random-effects model; risk of bias with QUADAS-2; PROSPERO: CRD420251103690. Five studies (1089 images) met criteria. Pooled sensitivity was 90.7% (95% CI, 85.6-94.1) and specificity 91.8% (95% CI, 86.0-95.3). CBCT-based models outperformed panoramic models, and transformer/hybrid architectures showed slightly higher accuracy than CNNs, though subgroup power was limited. Heterogeneity was moderate (I² ≈ 55-61%), plausibly related to variable diagnostic thresholds, mixed tooth types, and differing reference standards (orthodontists/endodontists/radiologists). Small evidence base (n = 5), small geographically limited datasets, and absence of true external validation restrict generalizability. With <10 studies, formal publication-bias testing was not feasible. AI shows high accuracy for ERR/ECR, particularly with CBCT and transformer/hybrid models, yet moderate heterogeneity and limited generalizability warrant cautious interpretation. AI should augment, not replace, clinician judgment within explainable, standardized workflows. Priorities include open multicenter annotated datasets, harmonized thresholds/protocols, and external validation to enable reliable clinical adoption.