[2503.13477] Periodontal Bone Loss Analysis via Keypoint Detection With Heuristic Post-Processing

View a PDF of the paper titled Analysis of Gingival Bone Loss via Keychain Detection with Post-Processing, by Ryan Banks and 6 other authors

View PDF HTML (beta)

a summary:This study proposes a deep learning framework and annotation methodology for automatic detection of periodontal bone landmarks and associated conditions, and staging. 192 Peripheral radiographs were collected and annotated with a stage-inappropriate methodology, describing clinically relevant landmarks regardless of disease presence or extent. We propose a post-processing marquee module that maps predicted keypoints to tooth boundaries using an additive instance segmentation model. An evaluation metric, relative percentage correct keypoints (PRCK), is proposed to capture keypoint performance in dental imaging fields. Four donor estimation models were adapted to our refinement of our keypoint problem. After processing, the fine localization was improved, raising the average PRCK^{0.05} by +0.028, but reducing the coarse performance of PRCK^{0.25} by -0.0523 and PRCK^{0.5} by -0.0345. Orientation estimation shows excellent performance for incremental segmentation when filtered with the Stage 1 object detection model. Gingival staging was adequately detected, with best intermediate and distal Dice scores of 0.508 and 0.489, while involvement of the fortis and median gingival space remained difficult due to rare positive samples. Implicit scalability with similar verification and external cluster performance. The annotation methodology allows for unobserved phase training with balanced representation across disease severity for some detection tasks. The PRCK metric provides a domain-specific alternative to general mode metrics, while the abstract post-processing unit constantly corrects implausible predictions with occasional catastrophic failure. The proposed framework demonstrates the feasibility of assessing clinically interpretable bone loss, with the potential to reduce diagnostic variability and clinicians’ workload.