1Department of Restorative and Prosthetic Dental Sciences, College of Dentistry, Dar Al Uloom University, Riyadh, Saudi Arabia 2General Dentist, Saudi Arabia
Background: Untreated canals are among the most consistent findings in teeth that fail after root canal treatment, and the second mesiobuccal (MB2) canal of the maxillary first molar is the most frequently implicated of them. Whether a missed canal is found before retreatment depends almost entirely on how an image is read, and reading is where human performance is weakest: intra- and inter-observer agreement for canal anatomy is modest, and in previously treated teeth the beam-hardening and streak artefacts generated by gutta-percha, posts and coronal restorations degrade precisely the region that must be inspected. Artificial intelligence (AI), and convolutional neural networks in particular, has been proposed as a second reader for this task.
Objectives: To evaluate systematically the diagnostic accuracy of AI models for the detection, localisation and segmentation of missed or unobturated root canal anatomy on cone-beam computed tomography (CBCT) and intraoral radiographs; to appraise the methodological quality of the primary studies; and to determine how much of the available evidence actually addresses the retreatment scenario rather than the untreated tooth.
Methods: A structured search of PubMed/MEDLINE, Scopus, Embase, Web of Science and the Cochrane Library was performed for studies published between January 2015 and September 2026, following PRISMA 2020. Eligible studies applied AI to the identification, localisation or segmentation of root canal anatomy — with emphasis on MB2 and other missed canals — on CBCT, periapical or panoramic images, and reported at least one quantitative performance metric against a defined reference standard. Study characteristics, model architecture, dataset size, validation strategy, reference standard and diagnostic metrics were extracted. Methodological quality was appraised with QUADAS-2 and reporting completeness against the CLAIM checklist.
Results: Reported performance for MB2 detection is high but heterogeneous. In the only study designed explicitly around the retreatment scenario — unobturated MB2 canals in already-obturated maxillary molars — a U-Net achieved sensitivity 0.80, specificity 1.00, positive predictive value 1.00 and accuracy 0.90, yet segmentation of the same canals was poor. Pooled estimates from a recent meta-analysis of four CBCT studies gave sensitivity 0.82 and specificity 0.74 (0.96 after correction for publication bias), with deep learning outperforming classical machine learning on every metric and heterogeneity exceeding 90%. Larger and more recent frameworks report AUC around 0.85 and, importantly, measurable improvement in the performance of junior clinicians when used as decision support. On two-dimensional periapical images the best models matched but did not surpass expert clinicians. Across the corpus, datasets are small, external validation is rare, reference standards are inconsistently defined, and only a minority of studies include previously treated teeth at all.
Conclusion: AI performs well enough on this task to be useful as a second opinion, and there is now direct evidence that it raises the performance of less experienced operators. What it does not yet have is the evidence base required for autonomous use: the field is dominated by single-centre, internally validated, retrospective models tested on curated images of untreated teeth. The clinically relevant question — can an algorithm find an unobturated canal in a metal-laden CBCT of a failing root filling and does finding it change the outcome — remains largely unanswered. Priorities are multicenter datasets that deliberately include retreatment cases, mandatory external validation, standardized reference standards and reader studies with outcome endpoints.
Keywords: Artificial Intelligence; Deep Learning; Endodontic Retreatment; Missed Canals; Second Mesiobuccal Canal; Cone-Beam Computed Tomography; Diagnostic Accuracy; Systematic Review
Salah A Yousief., et al. “Artificial Intelligence in Endodontic Retreatment: Detection of “Missed” Root Canal Anatomy”. EC Dental Science 25.11 (2026): 01-15.
© 2026 Salah A Yousief., et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Open Access by ECronicon is
licensed under a Creative Commons Attribution
4.0 International License
Based on a work at www.ecronicon.net
