Deep learning-based assessment of periodontitis progression using serial panoramic radiographs: a multicenter study incorporating indirect and direct grading approaches.
Authors
Affiliations (7)
Affiliations (7)
- Department of Periodontology, Seoul National University Bundang Hospital, Seongnam, Korea.
- Department of Oral and Maxillofacial Surgery, Wonju College of Medicine, Yonsei University, Wonju, Korea.
- Harvard John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA.
- Department of Periodontology, College of Dentistry and Institute of Oral Bioscience, Jeonbuk National University, Jeonju, Korea.
- Research Institute of Clinical Medicine of Jeonbuk National University-Biomedical Research Institute of Jeonbuk National University Hospital, Jeonju, Korea.
- Department of Periodontology, Seoul National University Bundang Hospital, Seongnam, Korea. [email protected].
- Research Institute of Clinical Medicine of Jeonbuk National University-Biomedical Research Institute of Jeonbuk National University Hospital, Jeonju, Korea. [email protected].
Abstract
This multicenter study aimed to evaluate the diagnostic performance and clinical feasibility of a deep learning (DL) model for assessing periodontitis progression using serial panoramic radiographs. A total of 7,106 panoramic images from 1,378 patients were retrospectively collected from 3 university hospitals. A YOLOv8-based DL model was trained separately for indirect (single-timepoint) and direct (serial) assessments to segment and classify alveolar bone levels. Diagnostic performance metrics, including accuracy, sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve (AUC), were evaluated for both the DL model and human evaluators. Comparative ROC analyses and assessment-time comparisons were conducted to evaluate discriminative ability and efficiency. The DL model demonstrated high diagnostic accuracy, particularly for indirect grading (accuracy, 89.6%; AUC, 0.913), and outperformed both interns and periodontal residents. Direct grading also performed well (accuracy, 79.5%; AUC, 0.854), although this performance was slightly lower due to the greater complexity of evaluating serial images. The DL model outperformed human evaluators (accuracy range, 38.0%-80.0%) while requiring significantly less assessment time (91.0±6.8 seconds vs. 143.7±33.6 seconds for interns). This multicenter study demonstrates that DL-based assessment of periodontitis progression using serial panoramic radiographs achieves high diagnostic accuracy (89.6% and 79.5% for indirect and direct assessment, respectively) while markedly improving time efficiency. This approach has strong potential as a standardized adjunctive diagnostic tool for periodontal clinical decision-making.