Periodontal bone loss analysis via keypoint detection with heuristic post-processing.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Dentistry

Background:

• Periodontal bone loss is a critical indicator of gum disease severity.
• Accurate detection and staging of periodontal bone loss are essential for effective treatment planning.
• Current diagnostic methods can be subjective and time-consuming.

Purpose of the Study:

• To develop a deep learning framework for automatic detection of periodontal bone loss landmarks, associated conditions, and staging.
• To introduce a novel annotation methodology for stage-agnostic training.
• To propose a new evaluation metric, Percentage of Relative Correct Keypoints (PRCK), for dental imaging.

Main Methods:

• Collected and annotated 192 periapical radiographs using a stage-agnostic methodology.
• Developed a heuristic post-processing module with an auxiliary instance segmentation model.
• Adapted and fine-tuned four pose estimation models for keypoint detection.
• Utilized the proposed PRCK metric for performance evaluation.

Main Results:

• The heuristic post-processing module improved fine-grained localization but impacted coarse performance.
• Periodontal staging achieved sufficient detection with Dice scores up to 0.508.
• Tasks like furcation involvement detection remain challenging due to limited positive samples.
• The framework demonstrated scalability with similar performance on validation and external datasets.

Conclusions:

• The developed annotation methodology enables stage-agnostic training with balanced disease severity representation.
• The PRCK metric offers a domain-specific evaluation for keypoint detection in periodontology.
• The deep learning framework shows feasibility for clinically interpretable periodontal bone loss assessment, potentially reducing diagnostic variability and workload.