Assessment of undersampling strategies for accelerated multi-shell diffusion MRI
Date
Authors
Editor(s)
Advisor
Supervisor
Co-Advisor
Co-Supervisor
Instructor
BUIR Usage Stats
views
downloads
Series
Abstract
Diffusion Magnetic Resonance Imaging (dMRI) is a non-invasive imaging technique that can probe the Brownian motion of water molecules within the neurite tissue. Measuring diffusion in the brain by densely sampling the q-space allows quantification of neural microstructure characteristics on a much smaller scale than the regular imaging resolution. Diffusion Tensor Imaging (DTI), which can resolve diffusion anisotropy and primary fiber orientation, is one of the most clinically adopted dMRI techniques. However, white matter (WM) voxels in the brain often contain crossing fibers and complex neurite structures, requiring more sophisticated dMRI techniques. With the goal of resolving multiple diffusion orientations and increasing angular resolution, multi-shell High Angular Resolution Diffusion Imaging (HARDI) samples the q-space densely on multiple spherical surfaces with radii determined by the b-values. This technique allows inferring neurite characteristics of complex fiber bundles, enabling the diagnosis of many neurodegenerative diseases. This thesis focuses on two multi-shell dMRI methods: Multi-Shell Multi-Tissue Constrained Spherical Deconvolution (MSMT-CSD), a model-free method, and Neurite Orientation Dispersion and Density Imaging (NODDI), a model-based method. These methods are used to resolve the Orientation Distribution Function (ODF) and Orientation Dispersion (OD) in the brain. However, the requirement of high number of q-space measurements, combined with the need to acquire images with reversed phase encoding (PE) directions for susceptibility artifact correction, causes prolonged acquisition times and reduces the clinical utility of multi-shell dMRI.
This thesis proposes several different undersampling strategies to accelerate multi-shell dMRI, and compares their performances based on the quality of estimated dMRI metrics and ODFs. The first undersampling approach is directly applied to the q-space using Electrostatic Energy Minimization (EEM) to produce 2- or 3-shell schemes with different combinations of gradient directions per shell. The second approach uses acquisitions with reversed PE directions for a single b0 volume only. These strategies are applied to achieve acceleration rates of R=2 and R=3 on dMRI data from 20 subjects. The results suggest that each diffusion metric prefers a different undersampling strategy. For R=2 case, 3-shell strategies perform better in terms of metric fidelity and ODF accuracy. Specifically, gradient tables with a lower variance in the number of q-space points between consecutive shells are preferable. For R=3 case, 2-shell strategies perform better and the strategies containing more gradient points on the outer shell are preferable. The metric maps produced from the undersampled data contain all the necessary microstructural information and preserve diagnostic properties. These analyses will be useful in designing disease and metric-specific multi-shell dMRI gradient tables to ease clinical applications and shorten the acquisition time with minimum loss of information.