(G-245) A deep learning brain injury model for efficient generation of pretraining samples for transfer learning

Introduction:: Physics-based finite element (FE) models of the human brain are extensively used to study the biomechanical mechanisms of traumatic brain injury (TBI). A major limitation, nevertheless, is that they demand enormous computational resources in terms of runtime and hardware. Thus, they are infeasible for routine applications. To dramatically improve efficiency while retaining high accuracy, several deep learning brain injury models have been introduced. With sufficient accuracy, a deep learning brain injury model can serve as an effective surrogate to the baseline FE model, thereby enabling large scale impact simulations, rapid concussion risk assessment on the field, and iterative design of head protective devices. As a result, the brain modeling community has recommended further integration of deep learning models into future TBI biomechanics research and practice (Ji et al. 2022). Nevertheless, a deep learning model usually requires a large amount of training samples, which still necessitates enormous computational resources. In this study, we explore whether a high accuracy transformer neural network (TNN) (Wu et al. 2022) already developed can be used to efficiently generate pretraining samples to allow finetuning a convolutional neural network (CNN) (Ghazi et al. 2021) via transfer learning. This could guide the development of a future CNN model to limit the amount of fresh training samples required for costly simulations with another or upgraded baseline FE model.


Materials and Methods:: A TNN was previously developed based on the anisotropic Worcester Head Injury Model (WHIM) V1.0 (Wu et al. 2022). The model has a high estimation accuracy across a wide range of real-world impacts (R²>0.99 for voxelwise whole-brain peak strains). To create pretraining samples, we used lab reconstructed and mouthguard measured head impacts to generate a new batch of augmented impacts (# ranging from 1 k to 5 k, at a step size of 1 k; N=5). We first verified that TNN-estimated brain strains remain highly accurate by comparing them with directly simulated responses for 100 randomly selected impacts. We then used the estimated samples (efficient, in minutes) to pretrain a CNN, which was further finetuned by using additional training samples generated from previous direct FE model simulations (days to months to generate; # of samples ranging from 250 to 5 k; N=7). Finally, we used an independent testing dataset (N=191) to evaluate CNN-estimation accuracy compared to directly simulated brain strains.


Results, Conclusions, and Discussions:: The TNN remained accurate for estimating the augmented impacts (R²>0.97 with RMSE ~0.01). Figure 1 reports the finetuned CNN estimation accuracy in terms of “success rate” (SR) based on the independent dataset, for different combinations of the numbers of finetuning (x-axis) and pretraining samples. The SR was defined as the number of impacts that were considered sufficiently accurate, where the linear regression slope, k, and Pearson correlation coefficient, r, between the CNN-estimated and directly simulated responses did not deviate from “the perfect score of 1.0” by more than either 0.1 or 0.05 (top and bottom panels, respectively). The two thresholds represented a relatively loose or a more stringent criterion for accuracy assessment, respectively.

For baseline training without pretraining, SR increased with the increase of the # of finetuning samples, as expected. However, it plateaued after N >=3k, regardless of the accuracy threshold. Pretraining samples improved SR, which were most effective when the # of finetuning samples were < 1k for accuracy threshold of 0.1, and < 2k for accuracy threshold of 0.05.


Conclusions: This study supports the use of the TNN brain injury model for efficient generation of pretraining samples so that to improve a finetuning CNN estimation accuracy via transfer learning. This could reduce the necessary number of fresh finetuning samples from direct and costly simulations from another or upgraded baseline FE model. This finding has practical importance to limit the computational resources required for developing a future deep learning brain injury model.




Acknowledgements (Optional): : NSF award under grant No. 2114697

References (Optional): : Ghazi K, Wu S, Zhao W, Ji S (2021) Instantaneous Whole-Brain Strain Estimation in Dynamic Head Impact. J Neurotrauma 38:1023–1035. https://doi.org/10.1089/neu.2020.7281

Ji S, Ghajari M, Mao H, et al (2022) Use of brain biomechanical models for monitoring impact exposure in contact sports. Ann Biomed Eng 50:1389–1408. https://doi.org/10.1007/s10439-022-02999-w

Wu S, Zhao W, Ji S (2022) Real-time dynamic simulation for highly accurate spatiotemporal brain deformation from impact. Comput Methods Appl Mech Eng 394:114913. https://doi.org/10.1016/J.CMA.2022.114913