Real-time integrated modeling of soft tissue deformation and stress based on deep learning

Area of Science:

• Computational mechanics
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Real-time simulation of soft tissue deformation and stress is vital for realistic surgical training.
• Current neural network models often neglect stress field modeling and struggle with multi-physics data imbalances.
• Existing methods lack the comprehensive physical field rendering needed for advanced surgical simulators.

Purpose of the Study:

• To develop a neural network-based framework for real-time multi-physics modeling of soft tissues, incorporating stress field prediction.
• To address data distribution challenges in multi-physics modeling using Z-Score normalization.
• To improve the accuracy and efficiency of surgical simulators by enabling real-time rendering of soft tissue physical properties.

Main Methods:

• A neural network approach compactly encodes the nonlinear relationship between boundary conditions and physical fields (deformation and stress).
• Z-Score normalization is employed to balance feature scales across different physical fields, preventing bias.
• The framework was validated on 3D models including biological tissues (liver, spleen, kidney) and a cantilever beam.

Main Results:

• The method achieves significant computational acceleration, with 1000x to 10,000x efficiency gains over traditional methods.
• Accuracy is maintained with only a minimal loss of approximately 1% compared to conventional approaches.
• Effective prediction of both displacement and stress distribution in soft tissues was demonstrated.

Conclusions:

• The proposed framework successfully integrates real-time deformation and stress field prediction for soft tissues.
• This advancement holds significant potential for enhancing the fidelity and capabilities of surgical simulators.
• The model offers a robust solution for multi-physics simulation challenges in biomedical applications.