Real-time simulation of viscoelastic tissue behavior with physics-guided deep learning.

Area of Science:

• Computational mechanics
• Biomedical engineering
• Machine learning

Background:

• Finite element methods (FEM) are computationally expensive for real-time soft tissue simulation in virtual reality (VR) surgical training.
• Existing data-driven methods often neglect physical laws, limiting their application to simpler hyperelastic models and ignoring time-dependent viscoelastic properties crucial for surgical scenarios.

Purpose of the Study:

• To develop a deep learning method for accurately predicting soft tissue displacement fields, incorporating viscoelastic properties for VR surgical training.
• To improve the computational efficiency and generalizability of soft tissue simulations in resource-constrained VR environments.

Main Methods:

• A deep learning model combining convolutional neural networks (CNNs) and long short-term memory (LSTM) layers was developed to predict spatiotemporal deformations.
• A physics-guided loss function, including a mass conservation law, was integrated to ensure physically plausible predictions and optimize model parameters.
• The model was trained on datasets generated using FEM simulations from a neurosurgery simulator.

Main Results:

• The physics-guided deep learning model demonstrated improved generalization for unseen simulation cases compared to conventional CNN models.
• Accuracy improvements ranged from 8% to 30% for unseen tissue deformations, depending on applied forces.
• The method successfully incorporated time-dependent viscoelastic properties, addressing limitations of prior data-driven approaches.

Conclusions:

• The proposed physics-guided deep learning approach enhances the accuracy and computational performance of soft tissue simulations for VR surgical training.
• This method bridges the gap between complex biomechanical modeling and real-time application requirements, improving the fidelity and utility of VR simulators.