A statistical shape modeling approach for predicting subject-specific human skull from head surface

Area of Science:

• Medical Imaging and Computational Anatomy
• Biomedical Engineering
• 3D Reconstruction

Background:

• Reconstructing internal skull structures from external head data is challenging.
• Existing methods lack subject-specific accuracy for complete skull generation.
• Accurate skull models are crucial for facial movement and simulation applications.

Purpose of the Study:

• To develop a novel, accurate method for predicting subject-specific 3D human skulls from head surface information.
• To utilize statistical shape modeling and Partial Least Squared Regression (PLSR) for head-to-skull prediction.
• To evaluate the accuracy and optimal parameters of the proposed prediction process.

Main Methods:

• A dataset of 209 CT scans was used to establish head-to-skull relationships.
• Feature points, distances, thickness, and volume descriptors were extracted for model learning.
• Hyperparameter tuning identified optimal feature points (2300), control points (1300), and PLSR components (4-8).
• Two learning strategies (point-to-thickness and distance-to-thickness, with/without volume descriptors) were evaluated using 10-fold cross-validation.

Main Results:

• The distance-to-thickness learning configuration demonstrated the best performance.
• Cross-validation yielded mean errors ranging from 2.46 ± 0.15 mm to 2.48 ± 0.27 mm.
• The best and worst predicted skulls showed mean Hausdorff distances of 2.09 ± 0.15 mm and 2.64 ± 0.26 mm, respectively.

Conclusions:

• A novel and accurate head-to-skull prediction process was successfully developed and validated.
• The method enables the first 3D subject-specific skull prediction from external head data.
• Future integration into real-time systems for facial animation and rehabilitation is planned.