Learning soft tissue deformation from incremental simulations.

Area of Science:

• Biomechanical modeling
• Computational anatomy
• Medical simulation

Background:

• Accurate biomechanical modeling of facial soft tissues is crucial for efficient orthognathic surgical planning.
• Traditional finite element method (FEM) simulations are time-consuming due to incremental steps, hindering clinical integration.
• Deep learning (DL) offers potential acceleration but often neglects temporal dynamics in incremental simulations.

Purpose of the Study:

• To investigate the efficacy of spatiotemporal incremental modeling for simulating facial soft tissue biomechanics.
• To develop and evaluate a DL-based approach that integrates spatial and temporal features for enhanced simulation accuracy and speed.

Main Methods:

• Implementation of a graph neural network (GNN) for spatiotemporal incremental modeling.
• Training DL networks on incremental FEM simulations from 17 orthognathic surgery patients.
• Synergizing spatial feature extraction with temporal aggregation for simulation.

Main Results:

• The spatiotemporal incremental method achieved a mean accuracy of 0.37 mm and a computation time of 1.52 s.
• A spatial-only incremental method resulted in 0.44 mm accuracy and 1.60 s computation time.
• A spatial-only single-step method showed 0.41 mm accuracy but a much faster 0.05 s computation time.

Conclusions:

• Spatiotemporal incremental modeling significantly reduced mean errors compared to spatial-only incremental methods, highlighting the value of temporal information.
• The developed DL approach effectively simulates soft tissue deformation, substantially decreasing simulation time versus FEM.
• This method shows promise for accelerating clinical surgical planning workflows.