PerioAID: An AI-Assisted Quality Assessment and Monitoring of Marginal Bone Levels in Bitewing Radiographs

Abstract Aim This study developed and evaluated PerioAID (Periodontal Artificial Intelligence-assisted Diagnostics), a framework for assessing bitewing radiograph quality, measuring marginal bone levels, monitoring radiographic bone loss, and assessing generalisability through external validation. Materials and Methods A total of 1000 sensor-based bitewing radiographs from 500 patients were annotated for cemento–enamel junction (CEJ) and marginal bone crests by two dental specialists. A You Only Look Once (YOLO) model was trained, tested, and externally validated on 100 phosphor plate-based radiographs. Performance was evaluated for image quality assessment, tooth detection, keypoint localisation, measurement accuracy, and longitudinal bone-level comparison. Results The model detected 95% of teeth and achieved 95% agreement for keypoint visibility. Measurements of CEJ-to-bone distances were comparable to interobserver variability and were maintained in the external validation cohort, demonstrating high diagnostic accuracy and generalisability. The model also identified radiographs unsuitable for bone-level assessment. In a longitudinal analysis, the framework enabled the comparison of marginal bone levels. Conclusions The PerioAID model accurately assessed image quality and marginal bone levels. Following further validation, it may serve as a clinical decision-support tool by flagging radiographs unsuitable for reliable bone-level assessment, facilitating early and longitudinal detection of marginal bone loss, and improving workflow efficiency.