An Automated Framework for Quantitative Alveolar Bone Loss Using Deep Learning-Based Landmark Detection.
Authors
Affiliations (5)
Affiliations (5)
- State Key Laboratory of Oral Diseases, West China School of Stomatology, National Clinical Research Center for Oral Diseases, Med-X Center for Manufacturing, Sichuan University, Chengdu 610064, China. Electronic address: [email protected].
- State Key Laboratory of Oral Diseases, West China School of Stomatology, National Clinical Research Center for Oral Diseases, Med-X Center for Manufacturing, Sichuan University, Chengdu 610064, China. Electronic address: [email protected].
- Department of Electrical and Computer Engineering, Zhejiang University; University of Illinois Urbana-Champaign Institute, Zhejiang University, Haining, P.R. China. Electronic address: [email protected].
- State Key Laboratory of Oral Diseases, West China School of Stomatology, National Clinical Research Center for Oral Diseases, Med-X Center for Manufacturing, Sichuan University, Chengdu 610064, China. Electronic address: [email protected].
- State Key Laboratory of Oral Diseases, West China School of Stomatology, National Clinical Research Center for Oral Diseases, Med-X Center for Manufacturing, Sichuan University, Chengdu 610064, China. Electronic address: [email protected].
Abstract
To develop and evaluate an automated framework for full-mouth quantification of radiographic alveolar bone loss (ABL) on panoramic radiographs by integrating deep-learning landmark detection with curve fitting. A total of 760 PANs (532/152/76) were annotated by two trained dentists, with third-expert adjudication. Three networks were evaluated: TransPose and HRNet for heatmap-based landmark localization, and YOLOv8 as a one-stage detector adapted for landmark regression. Landmark localization performance was evaluated using mean radial error (MRE) and success detection rate (SDR). Full-mouth and regional ABL percentages were derived via polynomial curve fitting, and agreement with manual measurements was quantified using intraclass correlation coefficients (ICC). Screening performance was evaluated at ABL thresholds of 15% and 33%. TransPose achieved the lowest MRE (12.50 px; IQR, 11.31-13.55) and the highest SDR (80.77 ± 5.79% at 15 px). YOLOv8 demonstrated the strongest agreement with manual measurements (ICC 0.633 for the maxilla; 0.771 for the mandible). In preliminary screening evaluation, YOLOv8 achieved sensitivity/specificity of 0.87/0.71 at the 15% cutoff and 0.57/0.98 at the 33% cutoff. Manual annotation required 14.2 min per image while automated analysis was completed within 12.3ms. The AI-based approach enables efficient quantification of ABL using PANs. TransPose achieved the highest localization accuracy, whereas YOLOv8 best aligned with clinical evaluations. This AI tool offers a rapid, objective alternative for ABL measurement on PANs, potentially supporting standardized screening in clinical practice.