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

Abstract

Abstract. Efficient deep learning brain injury models enable large-scale, strain-based investigations of traumatic brain injury (TBI). Here, we extend a previous model based on a generic 50th percentile adult male brain to a more individualized surrogate (applicable to male, female, and youth) for rapid estimation of brain deformation over impact duration. We utilize a previous training dataset (N = 1363) based on the randomly scaled anisotropic Worcester Head Injury Model (WHIM) V1.0 simulating random head impacts generated from real-world datasets. Brain–skull relative displacement was voxelized according to the three scaling factors along the anatomical axes at a temporal resolution of 1 ms. The earlier multitask convolutional neural network (CNN) architecture (i.e., splitting the impact duration into multiple time segments) was modified to take the scaling factors as three additional inputs for training with transfer learning. Based on ∼10% of the samples unseen from the training, the scaled CNN achieved an overall R2 of 0.96±0.05 and root mean squared error (RMSE) of 0.25±0.20 mm, respectively, at peak displacement magnitude with high efficiency (<1 s on laptop versus >30 min on a high-end cluster for direct model simulation). It also achieved an R2 of 0.83±0.16 and RMSE of 0.02±0.01 for the corresponding maximum principal strain (MPS). When using linear regression slope, k, and Pearson correlation coefficient, r, at time of peak displacement, the CNN achieved a success rate of 81.0% and 70.8% for displacement and MPS, respectively. This work could facilitate more individualized integration of repetitive head impacts in modeling and axonal injury model simulations in the future, important to fill a critical gap in a multiscale modeling framework to investigate white matter injury.