PhD Dissertation Defense
"Improve Efficiency and Accuracy of Brain Injury Estimation Through Deep Learning and Mesoscale Finite Element Modeling”
Nan Lin
Abstract: Concussion, a mild form of traumatic brain injury (mTBI), is common in contact sports and often underreported due to its subtle, transient symptoms and athletes’ reluctance to self-report. Although blood-based biomarkers and advanced imaging techniques are emerging, concussion diagnosis still heavily relies on subjective reporting, highlighting the need for objective and biomechanically-informed tools. Finite element (FE) human head models have been instrumental in understanding concussion mechanisms by estimating brain tissue responses such as strain. However, challenges remain in simulation efficiency, anatomical resolution, and the generalizability of models across individuals with different brain morphologies.
This dissertation addresses three critical gaps in concussion prediction by developing and evaluating multiple computational solutions across both FE and deep learning (DL) domains. First, the stability of simulated brain strain under varying kinematic input filters was assessed. Despite differences in peak angular velocity introduced by filtering, peak maximum principal strain (MPS) and its spatial distribution were found to remain relatively robust, supporting the consistency of strain-based injury metrics across studies using different preprocessing methods.
Second, a subject-specific convolutional neural network (CNN) was developed to account for individual brain size variations across three anatomical axes and achieved good prediction accuracy. However, training such models from scratch requires extensive data, often infeasible in practice. To address this, a method was developed to generate synthetic training data from limited real-world impact data. While synthetic data alone lacked the realism to fully train accurate CNNs, transfer learning using pretrained models significantly improved prediction accuracy, especially when real-world data were scarce.
Third, the feasibility and added value of mesoscale FE modeling were explored. A 2D mesoscale model demonstrated improvements in strain distribution prediction over coarsely meshed global models. Building on this, a novel 3D mesoscale model was developed to capture detailed strain and axonal strain distributions at the brain’s gray and white matter interface—a region frequently implicated in concussion pathology. The mesoscale model revealed sharp strain transitions at material boundaries not fully aligned with anatomical divisions, suggesting potential mechanical vulnerabilities.
Together, these studies contribute robust methodological advancements for improving concussion diagnosis and understanding brain injury biomechanics. They emphasize the value of anatomically informed modeling, the integration of synthetic data to overcome data scarcity in deep learning, and the potential of mesoscale models in uncovering localized injury mechanisms. This work lays the foundation for developing faster, individualized, and anatomically precise brain injury prediction tools in sports and clinical settings.
| Dissertation Advisor: | Committee Chair: | ||
|
Songbai Ji, PhD Professor Biomedical Engineering Worcester Polytechnic Institute
|
Karen Troy, PhD Professor & Department Head Biomedical Engineering Worcester Polytechnic Institute |
||
| Defense Committee: | |||
|
Zhenglun “Alan” Wei, PhD Assistant Professor Biomedical Engineering Worcester Polytechnic Institute |
Zhangxian Yuan, PhD Assistant Professor Aerospace Engineering Worcester Polytechnic Institute |
Lyndia Wu, PhD Assistant Professor Mechanical Engineering University of British Columbia |
|
For a zoom link, please email kharrison@wpi.edu