"APPLICATIONS OF DEEP LEARNING IN BRAIN INJURY BIOMECHANICS AND SPINE IMAGE REGISTRATION"
SHAOJU WU, Graduate Student
Abstract: Deep learning has been widely used in medical fields in recent years. It has shown to have huge potential to identify complicated patterns, such as biomarkers or symptoms, in large clinical datasets. However, its application in injury biomechanics, including traumatic brain injury (TBI), is very limited. In parallel, although deep learning techniques are now widely used in medical registration, including spine image registration in open spinal surgery, they are mostly used for volumetric images. To further expand the applications of deep learning in injury biomechanics and open spine surgery, we first developed deep learning pipelines that can be used to predict brain strain distribution of finite element (FE) models accurately and instantly for injury risk assessment in TBI. Then, we introduce a point cloud-based deep learning method to improve image registration efficiency and robustness in open spine surgery using pre-operative computed tomography (pCT) and intraoperative stereovision system (iSV).
In TBI studies, FE model of human head are important tools to understand the biomechanical mechanism of TBI. However, long computation time prevents the use of FE model in many real-world applications, such as concussion detection in the sports field. Although there are existing methods using pre-computation atlas and reduced order models to account for the computational cost of FE models, they either could not accurately predict the brain strains in complicated impacts collected from sport fields or could not efficiently generate dynamic responses of FE models, respectively. To address the computational cost of FE models, here we developed a convolutional neural network (CNN) to predict the regional brain strains for real-time, on field injury risk assessment in contact sport. Using our proposed CNN-based method, the regional brain strain, such as peak fiber strain of corpus callosum, could be accurately predicted based on the real-world sport-related head impact dataset. Furthermore, accurate estimation of dynamic brain strain is also critical to extract strain-rate based injury metrics for axonal injury analysis. Thus, to improve the efficiency for dynamic brain strain estimation, we advanced our CNN deep learning pipeline by using a transformer neural network (TNN) to predict five-dimensional (5D) brain displacement estimation. The estimated 5D displacement fields could be further transformed into a four-dimensional (4D) dynamic brain strain for injury risk assessment in TBI.
For the open spinal surgery, surgical image-guidance based on pre-operative images is becoming important, as it can help reduce injury risks of the spine during surgery, such as nerve impingement. Newly developed surgical image-guidance systems, which rely on the image registration between iSV and pCT images, have been proposed due to their advantages of low cost and radiation free compared to other image-guidance systems. However, the current image registration approach based on pCT-to-iSV may not be efficient. To further improve the spine registration robustness and efficiency, here we proposed a self-supervised learning approach with a point-cloud-based deep learning method, PointNet++, for vertebral segmentation, spine shape rectification, and 3D level-wise image registration. Based on the experimental results, our proposed deep learning-based registration pipeline was capable to further improve the speed and robustness in pCT-iSV image registration compared to other methods currently available.
In summary, this dissertation demonstrates that deep learning techniques can accurately and instantly estimate the static and dynamic tissue deformation for real-time, on field injury risk assessment in TBI. In addition, it also shows that a deep learning-based framework using point-cloud-based deep learning model is effective for cross-modality registration between CT and stereovision images in open spine surgery.
Dissertation Defense Committee: