Cephalometric landmark detection without X-rays combining coordinate regression and heatmap regression.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthodontics

Background:

• Current cephalometric landmark detection relies heavily on X-rays, posing radiation risks to patients.
• Advancements in convolutional neural networks offer potential for automated detection but are limited to X-ray modalities.

Purpose of the Study:

• To develop and evaluate a novel model for cephalometric landmark detection using only 2D facial profile images, thereby avoiding X-ray radiation.
• To assess the accuracy and reliability of the proposed non-X-ray method compared to existing cephalogram-based techniques.

Main Methods:

• A two-stage deep learning model was developed, utilizing high-resolution representation learning for initial landmark estimation from facial profiles.
• Refinement of landmark coordinates was performed by considering their spatial relationships and employing fully connected networks for enhanced accuracy.
• The model was trained and validated on a dataset of 2000 facial profile images from female patients.

Main Results:

• The proposed method achieved a Mean Radial Error (MRE) of 0.61 mm on test data.
• A high mean detection rate of 98.20% within a 2 mm tolerance was observed.
• Experimental results indicated performance on par with, or potentially exceeding, existing X-ray-based cephalometric methods.

Conclusions:

• The developed two-stage learning method accurately estimates cephalometric landmark positions using only facial profile images.
• This non-invasive approach demonstrates the potential to eliminate the requirement for X-rays in cephalometric landmark detection.
• The findings suggest a significant advancement in patient safety and diagnostic efficiency in cephalometrics.