Automated YOLO-Based Cephalometric Landmark Detection for ANB-Based Skeletal Classification: A Retrospective...

Area of Science:

• Orthodontics
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning for automated cephalometric landmark detection can enhance orthodontic diagnosis.
• AI accuracy's clinical relevance hinges on error propagation in measurements and classifications.
• Evaluating YOLO-based models for ANB-based skeletal classification agreement with expert diagnoses.

Purpose of the Study:

• To evaluate YOLO-based AI model configurations for cephalometric landmark detection.
• To quantify agreement between AI-derived and expert-derived ANB-based skeletal classifications.
• To assess the impact of model architecture, bounding-box size, dataset scale, and training epochs on accuracy.

Main Methods:

• Trained 12 YOLO-based models on lateral cephalograms, evaluating on an independent test set.
• Focused on Sella, Nasion, A-point, and B-point landmarks.
• Assessed localization accuracy (MRE, SDR) and downstream ANB classification concordance using statistical measures like Cohen's κ.

Main Results:

• The best model achieved 87.2% SDR@4 mm and 3.10±1.00 mm MRE.
• ANB-based skeletal classification showed 96.9% concordance with expert assessments (κ = 0.946).
• Bounding-box size significantly impacted localization accuracy, with smaller boxes yielding better results.

Conclusions:

• YOLO-based AI demonstrates promising diagnostic concordance for ANB-based skeletal classification.
• Prospective, multi-center validation is needed before clinical deployment.
• A confidence-aware workflow and careful bounding-box calibration are recommended for AI implementation.