Automated landmarking for palatal shape analysis using geometric deep learning.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Dental Morphology

Background:

• Accurate landmark identification on dental casts is crucial for clinical assessment and research.
• Manual landmarking is time-consuming and subject to inter-observer variability.
• Automated methods offer potential for increased efficiency and consistency.

Purpose of the Study:

• To develop and validate a geometric deep-learning network for automated placement of seven palatal landmarks.
• To assess the repeatability and accuracy of the automated landmarking system.
• To compare automated landmarking with manual methods in terms of palatal shape and size estimation.

Main Methods:

• A geometric deep-learning network was designed to process 3D point-cloud data from maxillary dental casts.
• The network was trained on 732 manually annotated casts and evaluated on 104 independent casts.
• Hierarchical feature learning was employed to predict the precise locations of seven key palatal landmarks.

Main Results:

• The automated system demonstrated excellent repeat-measurement reliability (<0.3 mm for all landmarks).
• Accuracy was promising, with 68-93% of landmarks correctly placed within 2 mm, varying by landmark.
• The method showed no significant difference in mean palatal shape estimation compared to manual methods and reduced measurement variability.

Conclusions:

• The developed deep-learning network provides a highly repeatable and accurate method for automated palatal landmarking.
• This technology has the potential to significantly streamline patient assessment and dental research.
• Visual quality control of landmark placement is recommended for optimal clinical application.