Related Experiment Video
Updated: Sep 23, 2025
05:30
Controlled Cortical Impact Model for Traumatic Brain Injury
Published on: August 5, 2014
28.9K
Real-time dynamic simulation for highly accurate spatiotemporal brain deformation from impact
Shaoju Wu1, Wei Zhao1, Songbai Ji1,2
1Department of Biomedical Engineering, Worcester Polytechnic Institute, Worcester, MA, United States of America.
Computer Methods in Applied Mechanics and Engineering
|May 16, 2022
Summary
Transformer neural networks (TNN) and convolutional neural networks (CNN) accurately estimate brain-skull displacement from impacts. TNN slightly outperforms CNN, enabling efficient real-time dynamic simulations for injury modeling.
Area of Science:
- Biomechanics
- Computational Neuroscience
- Machine Learning
Background:
- Real-time dynamic simulation of high-dimensional spatiotemporal data, particularly brain-skull impact, presents significant challenges.
- Accurate estimation of brain displacement and strain is crucial for understanding axonal injury mechanisms.
Purpose of the Study:
- To establish and compare Transformer Neural Network (TNN) and Convolutional Neural Network (CNN) models for estimating five-dimensional (5D) spatiotemporal brain-skull relative displacement and four-dimensional (4D) maximum principal strain resulting from impact.
- To evaluate the effectiveness of sequential training for these neural networks in achieving accurate, instantaneous estimations.
Main Methods:
- Developed and sequentially trained a TNN and a CNN to estimate 5D brain-skull displacement (4 mm spatial, 1 ms temporal resolution) over 60 ms duration using 5184 impact samples.
- Validated model accuracy using an independent testing dataset (N=314) and assessed performance across diverse real-world impact scenarios (dummy, helmet, football, soccer, car crash).
Main Results:
- Both TNN and CNN models achieved high accuracy in estimating displacement and 4D maximum principal strain (TNN: RMSE ~1.0%, R^2 >0.99; CNN: RMSE ~1.6%, R^2 >0.98).
- TNN demonstrated slightly superior accuracy over CNN.
- Models showed similar high accuracy across various real-world impact types, with average R^2 ~0.98 at peak displacement.
Conclusions:
- Sequential training enables accurate and instantaneous estimation of 5D brain-skull displacement and 4D strain, crucial for downstream multiscale axonal injury modeling.
- This study marks the first application of TNN in biomechanics, demonstrating its potential for real-time dynamic simulations in engineering.
- TNN offers a promising approach for advancing dynamic simulation capabilities across diverse scientific and engineering fields.
