Wu, Xuehai. White matter hybrid computational simulations and 3D convolutional neural networks framework. Retrieved from https://doi.org/doi:10.7282/t3-zrwp-qy61
DescriptionThe finite element methods were widely utilized in White Matter (WM) modeling to predict and analyze brain axon injury. The brain axon injury was caused by both the strain magnitude and the strain rate. The purpose of this study is to combine the three fundamental requirements for simulation of WM, which are the axon/neuroglia composite anisotropy, the axonal trace orientations, and dynamic response in the frequency domain. A triphasic (axon, myelin, glia), 2D representative elemental volumes (REVs) was developed to simulate the influence of each phase's intrinsic viscoelastic property and volume fraction. A sensitivity analysis has been applied to the effective transverse modulus to determine which model parameters are essential to the estimated REV-based MRE metrics. In the 3D orthotropic REV model, the information of axonal traces is exploited in this study to specify the material orientation in finite element modeling. The material orientation specification will consider 3D axonal trace position data, which can offer axonal traces directional and locational information. Numerous REVs are integrated as the elements of the WM model based on typical volume fractions (VFs) calculated by the relative distance between the elements and axonal traces. A deep 3D convolution neural network (CNN) algorithm combined with the 3D anisotropic REV model was employed to predict the WM's anisotropic material properties. The architecture information encoded in the voxelated location is used as input data and is consequently incorporated into a 3D CNN model that cross-references the REVs' material properties (output data). The output data (REVs' material properties) is calculated in parallel using an in-house developed finite element method, which models REV samples of axon-myelin-glia composites. This novel combination of the CNN-REV method dramatically reduced the computation time compared with directly using finite element methods.