A Shape and Size-Scaled Deep Learning Brain Injury Model for Near Real-Time Dynamic Impact Simulation

Area of Science:

• Biomechanics
• Computational modeling
• Neuroscience

Background:

• Traumatic brain injury (TBI) research requires efficient models for large-scale, strain-based investigations.
• Previous models were limited to generic brain surrogates, hindering individualized analysis.
• Accurate estimation of brain deformation and strain is vital for understanding TBI mechanisms.

Purpose of the Study:

• To extend an existing deep learning model for traumatic brain injury (TBI) to an individualized surrogate applicable to males, females, and youth.
• To enable rapid estimation of brain deformation and strain over impact duration.
• To improve the integration of repetitive head impacts in multiscale modeling frameworks.

Main Methods:

• Utilized a training dataset (N=1363) from a scaled anisotropic Worcester Head Injury Model (WHIM) V1.0 simulating head impacts.
• Voxelized brain-skull relative displacement based on three anatomical scaling factors at 1ms temporal resolution.
• Modified a multitask convolutional neural network (CNN) architecture with transfer learning, incorporating scaling factors as additional inputs.

Main Results:

• The scaled CNN achieved high accuracy for peak displacement magnitude (R2=0.96±0.05, RMSE=0.25±0.20mm) and maximum principal strain (MPS) (R2=0.83±0.16, RMSE=0.02±0.01).
• Achieved high efficiency, with predictions in <1s compared to >30min for direct simulations.
• Demonstrated success rates of 81.0% for displacement and 70.8% for MPS at peak impact.

Conclusions:

• The developed individualized deep learning model enables efficient and rapid estimation of brain deformation and strain.
• This approach facilitates more personalized modeling of head impacts and axonal injury.
• The work addresses a critical gap in multiscale modeling for white matter injury research.