Improved Simulation, Bayesian Estimation, Phase Activation, and Non-Cartesian Reconstruction in FMRI
Date of Award
Spring 4-28-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Mathematical and Statistical Sciences
First Advisor
Daniel Rowe
Second Advisor
Cheng-Han Yu
Third Advisor
Mehdi Maadooliat
Fourth Advisor
Tugan Muftuler
Abstract
In fMRI it is desirable to obtain images with high contrast and low noise at a high spatio-temporal resolution. Doing so allows for a more reliable assessment of capturing brain dynamics in higher detail. Many techniques have been employed and tested to collect more images per unit of time and/or increase spatial resolution, while maintaining a high quality image. This has been done using in-plane and through-plane acceleration with promising results (SENSE, GRAPPA, CAIPIRINHA, CAIPIVAT). Following a description of work done towards a unified fMRI simulation software package, this dissertation will present work towards three methods to enhance fMRI time series image quality using simulated data as the benchmark. The simulation tool presented is titled Simulation and Harmonic Analysis of k-Space Readout (SHAKER) and is designed to simulate fMRI time series for MR scientists. At present, most researchers involved in fMRI simulate their own data in-house and do not necessarily do so in a way that is representative of the machine. SHAKER aims to provide software that allows for fast simulations to be done without the need for advanced understanding of how an MRI machine works. The images and time series simulated by SHAKER will be used to demonstrate the efficacy of the three new techniques presented that will improve fMRI images. The first method makes use of the first three k-space arrays in fMRI time series that are often discarded due to having a higher signal than the remaining images. These brighter images will be used to assess hyperparameters for prior distributions that will be combined with distributionally accurate likelihood images from the steady-state time series to form posterior images that have higher signal and contrast, with lower noise. The second method focuses on phase-only activation in complex-valued fMRI time series. The phase half of the data is often discarded, and only the magnitude is studied. We will show that the phase part of the data contains biological information, in particular task-related signal change, that has exciting physiological implications. The third method introduced in this dissertation will operate in a radial k-space where instead of sampling on a Cartesian grid, points are collected on spokes that each pass through the center of k-space. This method of sampling k-space has many proven benefits, but is not without it's challenges- in particular, reconstructing the k-space arrays back into images. A few, fully-sampled radial k-space arrays can be used to assess hyperparameters for prior distributions that can be combined with subsampled likelihood images from the fMRI time series to form posterior images. It is expected that these images will have higher signal and contrast, lower noise, and measured at significantly higher temporal resolutions than conventional subsampling techniques offer. The software tool SHAKER and the three techniques described are novel and significant contributions to fMRI analysis.