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

TL;DR

This work tackles automatic periodontal bone loss assessment from periapical radiographs by formulating joint object-detection, keypoint-detection, and instance-segmentation tasks within a unified framework. It introduces a stage-agnostic annotation protocol to balance data across disease severities, a novel Percentage of Relative Correct Keypoints ($PRCK$) metric normalized by the average tooth diagonal length $L$, and a heuristic post-processing module that realigns keypoints to tooth boundaries using an auxiliary segmentation model. Empirical results show strong performance for tooth detection and keypoint localization on common morphologies, with quantitative gains in strict keypoint accuracy from post-processing but some reductions at coarser thresholds, and reasonable periodontal staging for mesial/distal sites though furcation involvement and widened periodontal ligament space remain challenging due to data scarcity. The approach contributes to clinically interpretable, scalable periodontal assessment with potential to reduce diagnostic variability and clinician workload, while highlighting the need for further anatomical constraints integration and broader representation of underrepresented conditions. Key contributions include the dataset annotation protocol, the $PRCK$ domain-specific metric, and the segmentation-guided post-processing pipeline, all validated on external data to support scalability and generalisability. $L$ denotes the average tooth-diagonal length used for normalization, and $d_{thresh}\in\{0.5,0.25,0.05\}$ defines evaluation strictness in the $PRCK$ metric.

Abstract

This study proposes a deep learning framework and annotation methodology for the automatic detection of periodontal bone loss landmarks, associated conditions, and staging. 192 periapical radiographs were collected and annotated with a stage agnostic methodology, labelling clinically relevant landmarks regardless of disease presence or extent. We propose a heuristic post-processing module that aligns predicted keypoints to tooth boundaries using an auxiliary instance segmentation model. An evaluation metric, Percentage of Relative Correct Keypoints (PRCK), is proposed to capture keypoint performance in dental imaging domains. Four donor pose estimation models were adapted with fine-tuning for our keypoint problem. Post-processing improved fine-grained localisation, raising average PRCK^{0.05} by +0.028, but reduced coarse performance for PRCK^{0.25} by -0.0523 and PRCK^{0.5} by -0.0345. Orientation estimation shows excellent performance for auxiliary segmentation when filtered with either stage 1 object detection model. Periodontal staging was detected sufficiently, with the best mesial and distal Dice scores of 0.508 and 0.489, while furcation involvement and widened periodontal ligament space tasks remained challenging due to scarce positive samples. Scalability is implied with similar validation and external set performance. The annotation methodology enables stage agnostic training with balanced representation across disease severities for some detection tasks. The PRCK metric provides a domain-specific alternative to generic pose metrics, while the heuristic post-processing module consistently corrected implausible predictions with occasional catastrophic failures. The proposed framework demonstrates the feasibility of clinically interpretable periodontal bone loss assessment, with potential to reduce diagnostic variability and clinician workload.

Figures (9)

• Figure 1: Three images containing example annotations of the collected keypoints and rotating bounding boxes.
• Figure 2: (a): Bar plot showing the counts of bounding box class instances of our baseline dataset before processing and after cleaning. (b): Bar plot showing the counts of keypoint instances of our baseline dataset before processing and after cleaning.
• Figure 3: Images containing predicted segmentation mask overlays, for Image 1, where (a) is before NMM and (b) is after NMM.
• Figure 4: Handmade example diagrams, with synthetic data, depicting each stage of the post-processing module overlaid on Image104.
• Figure 5: Training/inference loops for the two-stage top-down pipeline with post-processing, and the end-to-end YOLOv8 variant.