Development and Validation of AI System for Tooth Detection and Diagnosis in Dental Radiographs.
Authors
Affiliations (6)
Affiliations (6)
- Department of Oral and Maxillofacial Surgery, Radboud University Medical Center, Nijmegen, The Netherlands. Electronic address: [email protected].
- AID s.r.o., Bratislava, Slovakia.
- Department of Oral, Maxillofacial, and Plastic Facial Surgery, University Hospital Münster, Münster, Germany.
- Department of Oral and Maxillofacial Surgery, Radboud University Medical Center, Nijmegen, The Netherlands.
- Department of Oral and Maxillofacial Surgery, Erasmus University Medical Center, Rotterdam, The Netherlands.
- Department of Dentistry, Radboud University Medical Center, Nijmegen, The Netherlands.
Abstract
To develop and validate an AI-automated system for dental charting that accounts for multiple radiograph types and younger patients. A total of 3705 dental radiographs were collected in Slovakia and Egypt between 2023 and 2024, including orthopantomograms (OPGs) and intraoral radiographs (e.g. bitewings). Complete teeth were manually annotated with bounding boxes and tooth numbers by 2 calibrated annotators. Subsequently, teeth were assessed by at least 3 trained clinicians (10 total, 2-13 years of clinical experience) for the presence of ten dental findings (caries, crown, filling, implant, pontic, periapical lesion, primary tooth, retained root, root canal treatment, unerupted tooth). Assessments were aggregated using the Dawid-Skene model. The AI system comprised 3 deep learning stages for modality classification (OPG or intraoral radiograph; EfficientNetV2), tooth detection per modality (RTMDet), and dental finding classification (EfficientNetV2) and was evaluated with 5-fold cross-validation on 400 held-out radiographs against the clinicians and 3 independent dentists. Tooth detection was highly effective (F1-score: OPG = 0.99, intraoral = 0.98; tooth number accuracy: OPG = 0.98, intraoral = 0.96) with decreased effectiveness for primary teeth. Dental finding classification saw mixed effectiveness (F1-score: 0.52 to 0.99) with a lower effectiveness for disease-related findings. The system was more accurate than the dentists for caries, crowns, fillings, primary teeth, root canal treatments, and unerupted teeth. The AI system outperformed dentists for common findings, but more radiographs and annotations are required to effectively interpret rare conditions. Overall, it can support radiographic diagnosis and speed up dental charting for most patients. Interpretation of dental radiographs requires clinical experience and remains challenging. An effective AI system was developed for tooth detection and diagnosis in several types of dental radiographs of young and adult patients. Its findings can enhance the clinical workflow, facilitate dentist-patient communication, and improve diagnostic consistency for prevalent findings.