Algorithms at the Apex: A Scoping Review of Artificial Intelligence across the Endodontic Workflow, from Diagnosis to Prognosis and Education

Abstract

Introduction: Artificial intelligence (AI), particularly convolutional and artificial neural networks, is being applied to image interpretation, decision support and outcome prediction in endodontics. Since the scoping review by Asgary in 2024, the literature has expanded to include large language models, microrobotics and the first meta-analysis in the field. This scoping review aimed to map the current evidence on AI in endodontics across the clinical workflow and to identify gaps that limit clinical translation. Materials and Methods: The review followed the Arksey and O’Malley framework, Joanna Briggs Institute guidance and the PRISMA-ScR checklist. Sources published between 2021 and 2026 addressing any application of AI in endodontics or closely related restorative fields were eligible. Data were charted on source type, scope, AI approach, clinical task and reported performance, and synthesised thematically. Results: Twenty sources were included, 16 (80%) published from 2024 onward. Nineteen were secondary sources: ten narrative reviews, five structured literature reviews, one scoping review, one umbrella review, one systematic review with meta-analysis and one perspective. Only one was a primary study. AI applications clustered in diagnosis (periapical lesions, vertical root fractures, caries, root canal morphology), treatment (working length, case difficulty, retreatment decisions), prognosis (postoperative pain, retreatment success) and education. Reported accuracy exceeded 90% for periapical lesion detection and reached 95.6% for postoperative pain prediction. The only meta-analysis found higher odds of correct working-length determination with AI than with expert assessment, but it was based on five laboratory studies and the certainty of evidence was low. Large language models answered 25% to 72% of restorative and endodontic examination questions correctly, and performed worst on endodontics. Recurrent barriers were small and unrepresentative datasets, limited external validation, poor interpretability, cost, and ethical and privacy concerns. Conclusions: AI shows high reported accuracy for image-based endodontic tasks, but the evidence base is dominated by reviews that repeatedly summarise a limited pool of internally validated, largely laboratory-based studies. Prospective, multicentre clinical validation, standardised reporting and ethical and regulatory frameworks are needed before routine clinical use.