BOLD fMRI (Blood-Oxygenation Level Dependent functional Magnetic Resonance Imaging) measures the haemodynamic correlates of brain function, with research and clinical applications. However, fMRI is limited by relatively weak signal, and large, complex noise sources. A variety of preprocessing algorithms have been developed to remove artifacts and improve signal detection, but there is no literature consensus on optimal preprocessing strategies. Furthermore, it is not well understood how fMRI experimental design choices interact with preprocessing steps.
This thesis develops a statistical framework for selecting the set of preprocessing choices (“pipelines”), using data-driven metrics of (R) reproducibility of brain maps, and (P) prediction of experimental stimuli. These metrics were used to evaluate standard pipeline steps on data from young healthy subjects, who performed a set of brief tasks in an fMRI cognitive assessment battery. It is shown that (1) preprocessing choices have significant, consistent effects on the detection of brain networks in fMRI. However, (2) optimizing pipelines on a subject- and task-specific basis, compared to the standard fMRI approach of applying a single fixed set of preprocessing choices, improves (P, R) and independent test measures of between-subject activation overlap. This indicates that signal detection in standard fMRI may be limited by sub-optimal pipeline choices.
Even after optimizing standard pipeline choices, physiological noise is a major confound in fMRI analysis; this includes BOLD signal changes due to respiration and pulsatile blood flow. As a potential solution, the PHYCAA (PHYsiological correction using Canonical Autocorrelation Analysis) algorithm is developed. This multivariate, data-driven model estimates physiological noise, without respiratory and cardiac measurements. The estimated noise has a spatial distribution consistent with non-neuronal tissues, and its dimensionality is correlated with cardiac and respiratory variability. Removing this physiological noise increases (P, R) of analysis results. The PHYCAA model provides novel information about the structure of physiological noise in fMRI, and a principled method of removing physiological artifact.
The results of this thesis were obtained using data from a prototype fMRI cognitive assessment battery, designed for clinical use. The datasets involve brief scanning sessions with complex cognitive tasks. These findings are therefore relevant for clinical implementation of fMRI.
| Identifer | oai:union.ndltd.org:TORONTO/oai:tspace.library.utoronto.ca:1807/43518 |
| Date | 08 January 2014 |
| Creators | Churchill, Nathan William |
| Contributors | Strother, Stephen |
| Source Sets | University of Toronto |
| Language | en_ca |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0039 seconds