Three-Dimensional Craniofacial Landmark Detection in Series of CT Slices Using Multi-Phased Regression Networks

Area of Science:

• Medical imaging and computer vision
• Anthropometry and craniofacial morphology
• Artificial intelligence in healthcare

Background:

• Accurate geometrical assessments of human skulls rely on identifying anatomical landmarks.
• Automated detection of these landmarks promises substantial medical and anthropological applications.
• Current methods may lack the precision and efficiency required for complex 3D analyses.

Purpose of the Study:

• To develop an automated system for precise 3D coordinate prediction of craniofacial landmarks.
• To leverage multi-phased deep learning networks for enhanced landmark detection accuracy.
• To validate the system's performance against human expert measurements.

Main Methods:

• Utilized computed tomography (CT) images from a public database, reconstructed into 3D objects.
• Developed a three-phased regression deep learning network trained on 90 datasets.
• Evaluated the system on 30 independent testing datasets to measure 3D error.

Main Results:

• The system achieved a significant reduction in 3D error across three phases: 11.60 px, 4.66 px, and finally 2.88 px.
• The final 3D error of 2.88 px is comparable to the landmark placement variability between experienced practitioners.
• The multi-phased approach progressively narrowed down landmark detection areas, optimizing accuracy.

Conclusions:

• The proposed multi-phased deep learning system effectively predicts 3D craniofacial landmark coordinates with high accuracy.
• This automated method offers a viable solution for landmark detection challenges, considering computational constraints.
• The findings support the potential of AI in advancing craniofacial analysis for medical and anthropological research.