Keypoint-based detection of periodontal radiographic bone loss on bitewing radiographs using YOLO11-pose: a clinician-AI agreement study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Radiographic assessment of periodontal bone loss on bitewing radiographs is a manual, interpretive task subject to inter- and intra-observer variability. This study aimed to develop and internally validate a YOLO11-pose model for automated localization of the cemento-enamel junction (CEJ) and alveolar bone crest (ABC) and to assess the agreement of its CEJ-ABC distance measurements with a multi-expert consensus reference. Methods A total of 927 bitewing radiographs (6,276 interproximal sites) from 729 patients were analyzed. The reference standard was established by simultaneous consensus among three specialists (one periodontologist and two oral and maxillofacial radiologists). Data were partitioned at the patient level using a 70:15:15 training-validation-test split. A YOLO11-pose model was trained to detect two keypoints (CEJ, ABC) per site, with the operating threshold selected on the validation set. Detection performance, keypoint localization error, and clinician-AI agreement were evaluated on a held-out internal test set. A YOLOv8-pose baseline was trained for comparison using identical partitions. Results On the held-out internal test set (141 radiographs; 882 interproximal sites), the model attained a precision of 0.672, recall of 0.778, and F1-score of 0.721. Among the 686 successfully detected and matched sites, the mean CEJ-ABC distance error was 4.05 pixels, and clinician-AI agreement was good (ICC(2,1) = 0.850, 95% CI 0.799–0.887; Pearson r = 0.860; Bland-Altman mean bias, − 1.28 pixels). Patient-level averaging yielded higher agreement (ICC(2,1) = 0.941). In the controlled comparison, YOLO11-pose showed a modestly higher F1-score (0.721 vs. 0.705) and lower distance error (4.05 vs. 4.71 pixels) than the YOLOv8-pose baseline, with essentially equivalent precision; this internal benchmark does not establish general architectural superiority. Conclusions Model-derived CEJ-ABC distances showed good agreement with the consensus reference, supporting the potential of keypoint-based deep learning as a second-reader measurement aid for periodontal radiographic assessment. External multicenter validation and prospective clinical evaluation are required before implementation in clinical workflows.