Dissertation/Thesis Abstract

Diffusion MRI white matter tractography: Estimation of multiple fibers per voxel using Independent Component Analysis
by Wong, Chi Wah (Alec), Ph.D., University of Southern California, 2009, 134; 3368725
Abstract (Summary)

Brownian motion is a phenomenon describing the random movement of particles. Water molecules, the most abundant particle inside the body, diffuse in all directions with equal probability in free space. In the presence of organized structure, diffusion is restricted. Diffusion Tensor Imaging (DTI) was developed to model water diffusion in tissues. In the human brain, myelinated neuronal fiber bundles hinder water diffusion. By knowing the amount of diffusion along different directions, the underlying fiber bundle directions can be inferred.

DTI is implemented by adding pulsed magnetic field gradients to a T 2-weighted spin echo pulse sequence. A diffusion signal is obtained for each voxel by combining data using different gradient directions. Apparent Diffusion Coefficient (ADC) profile is calculated to show the amount of diffusion in different directions. A rank-2 tensor (ellipsoid) is then used to model the diffusion profile. From the tensor, the direction of the fiber bundle can be estimated. This technique is generally regarded as Principal Component Analysis (PCA).

However, a rank-2 tensor model fails to account for the possibility of multiple fibers in a single voxel. When multiple fiber bundle directions exist within a voxel, the PCA estimated orientation lies somewhere in-between the multiple fiber directions, and is an incomplete and incorrect representation of the underlying fibers in the voxel.

A major problem of existing multiple fiber reconstruction methods is that they require High Angular Resolution Diffusion Imaging (HARDI) which usually takes 60 to 100+ gradient directions with multiple averages, requiring at least 15 minutes per scan.

In this thesis, a novel method is proposed using Independent Component Analysis (ICA) to estimate the orientations of multiple fiber bundles in a single voxel. Diffusion data used with this technique is obtained using clinical pulse sequences and are acquired in around 3-7 minutes. The novelty of the approach is to use the non-Gaussian property of diffusion signals arising from individual fiber directions. White matter neuronal fiber bundles are in orders of millimeters, which are larger than a voxel, is utilized. It is very likely that a fiber bundle will span the width of several voxels and this neighborhood information is used to estimate multiple fibers.

First, the pre-requisites of ICA were examined. By calculating the negentropy of the diffusion data from voxels known to contain only a single fiber direction (corpus callosum) and bootstrapping, we found that the diffusion signal caused by single fiber direction is non-Gaussian.

Monte Carlo simulation studies were carried out to estimate the accuracy of recovering the number of fiber directions per voxel, as well as the accuracy of individual recovered fiber directions. Results showed that the average absolute error angle was less than 5 degrees with inter-fiber angle of above 50 degrees in the case of two fiber directions per voxel at SNR=30 and SNR=45.

By incorporating multiple fiber streamline tractography, the method was validated on synthetic phantom data, which revealed significant improvement over both DTI and an alternative multi-compartment approach. The approach was then applied to experimental human data.

Indexing (document details)
Advisor: Singh, Manbir
Commitee: Leahy, Richard, Nayak, Krishna, Shung, K. Kirk
School: University of Southern California
Department: Biomedical Engineering
School Location: United States -- California
Source: DAI-B 70/08, Dissertation Abstracts International
Subjects: Biomedical engineering, Electrical engineering
Keywords: Diffusion MRI, Diffusion tensor imaging, Fiber crossing, Independent component analysis, Magnetic resonance imaging, Multiple fibers
Publication Number: 3368725
ISBN: 9781109295214
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy