Enhanced head-skull shape learning using statistical modeling and topological features.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Anatomy

Background:

• Skull prediction from head data is crucial for cost-effective facial disorder treatments.
• Previous methods using full head-to-skull relationships had limitations due to locally varying skull thickness, especially in the facial region.

Purpose of the Study:

• To enhance head-to-skull prediction accuracy by integrating local topological features of the head and skull.
• To improve skull shape prediction for better therapeutic solutions for facial disorders.

Main Methods:

• Sampled feature points from 329 computed tomography (CT) head and skull models.
• Classified feature points into back and facial topologies.
• Trained partial least square regression (PLSR) models separately for each topology with hyperparameter tuning.

Main Results:

• Topology-based learning significantly improved prediction accuracy compared to non-topology-based methods.
• Mean error reduction of 36.96% for skull shapes and 14.17% for skull models after tenfold cross-validation.
• Substantial error improvements noted in facial skull regions (4.98%), muscle attachment regions (11.71%), and back skull regions (25.74%).

Conclusions:

• Detailed local topological information of the head-skull shape is essential for accurate skull prediction.
• The enhanced learning strategy, considering local features, improves skull prediction accuracy for facial disorder applications.
• This method can enhance clinical decision support systems and warrants further investigation into geometric deep learning.