Mild Traumatic Brain Injury is a major health concern in the United States and around the world and is especially common in contact sports and is difficult to reliably diagnose as it is often diagnosed in a symptom-based fashion. This is problematic since the symptoms may take time to develop and the recognition of the symptoms could introduce some subjective bias. In addition, athletes are less likely to recognize, appreciate the significance of, or disclose symptoms in a competitive atmosphere such as that of contact sports. This is a contributing factor to the difficulty of diagnosis. Undiagnosed mTBI can cause more serious health complications such as neurodegenerative diseases. Hence, there is a significant need to reliably predict the risk of concussion, and prevent concussion using preventative equipment. One method that holds a great potential for such scenarios is FE modeling. Yet FE models are computationally expensive, making them infeasible in a side-line scenario. As a result there has been a recent shift to pre-computation based techniques to bypass the time consuming FE simulations. Yet, referring to the pre-simulated database to exploit the full potential of such methods is a challenge as a result of the complex nature of impact profiles. Here, we addressed this challenge by using deep learning based approaches, which are well-suited for modeling such complex scenarios with constant boundary conditions.
Here, we first identified a gap in the literature about the implications of the used software packages for FE simulation as a potential point of discrepancy in FE based TBI research. Hence, we established a bridge between Abaqus and LS-DYNA, which are two of the most widely-used software platforms currently used in FE modeling. We identified the differences between the two packages along the way to convert our Worcester Head Injury Model with material anisotropy, which was originally developed in Abaqus, into LS-DYNA format. We used the most reliable element type in Abaqus (C3D8I) as a benchmark, converted the WHIM with 17 C3D8I elements through a series of steps into an isotropic version with C3D8R elements that could be directly translated into LS-DYNA format, and translated the model into LS-DYNA format without any alterations. Then we compared the model with all the appropriate LS-DYNA model configurations to the Abaqus model. Then we identified the LS-DYNA configurations that perform the most similarly with Abaqus hence stablishing a bridge between the two. Further in the aims sections, with the assumption of limited time and computational power, we bridged the gap between FE based TBI research and utilization of such methods in real-world scenarios. We developed and assessed different machine learning based approaches with the goal of making different aspects of FE based injury assessment real-time. In the first aim, using a deep learning approach, we bypass FE based model simulations entirely and obtain the entire brain strain pattern directly from the impact profile. In the second aim, we developed a new, strain-based injury metric using an inverse approach to map brain strains into a simple kinematic profile. The advantage of this approach is that unlike all other available kinematic based injury metrics, it accounts for impact directionality. Finally in the third aim, we use our developed methods in a real-world scenario by investigating the effectiveness of a number of helmets based on the generated strains in the brain.
Please contact email@example.com for a zoom link to this event.