Evaluation of the Diagnostic Performance of an Implant Marginal Bone Loss Measurement Method Based on Deep Learning.

OBJECTIVES: To evaluate the diagnostic performance of the YOLOv8-pose model for automated measurement of mesial and distal peri-implant marginal bone loss on periapical radiographs. METHODS: In this retrospective study, 162 images were divided into training (n=144) and validation (n=18) sets; 46 images formed an independent test set. Six anatomical key points were annotated and verified as the gold standard. The model was tested at thresholds of bone loss ≥0.5 mm, ≥1.0 mm, and ≥2.0 mm. F1-score was additionally calculated for each threshold. RESULTS: At the 0.5 mm threshold, the model achieved 95.35% sensitivity, 100% specificity, 100% positive predictive value (PPV), and 90.7% mesial-distal diagnostic consistency. At the 1.0 mm threshold, sensitivity was 91.67%, specificity 82.14%, PPV 73.33%, and consistency 58.14%. At the 2.0 mm threshold, sensitivity dropped to 30.43%, while specificity remained high at 95.97%, with 53.85% PPV and 72.1% consistency. Median key point localization errors ranged from 0.76 mm to 1.05 mm, with no statistically significant difference among groups (P>0.05). CONCLUSIONS: The YOLOv8-pose model shows excellent diagnostic performance, especially for early mild bone loss. This artificial intelligence tool serves as an auxiliary early-screening tool rather than a comprehensive diagnostic tool covering all severities of bone loss; it can assist clinicians in standardized, efficient, and early screening assessment of peri-implant diseases. ADVANCES IN KNOWLEDGE: This study validates a keypoint-based deep learning approach for automated and consistent measurement of peri-implant marginal bone loss on periapical radiographs, supporting early objective screening-level assessment of radiographic peri-implant marginal bone-loss changes for clinician reference.