Methodologies to enhance the reliability of functional MRI analysis

Kihwan Han, Purdue University

Abstract

Contributing to the growing popularity of functional magnetic resonance imaging (fMRI) as a noninvasive neuroimaging modality, researchers have advanced the credibility of fMRI analysis through algorithm development and hardware enhancement. In this context, this dissertation presents two methodologies to improve the interpretability of fMRI analysis. The first project addresses a practical issue for large-scale studies, arising from the timing of hardware upgrades, with a particular focus on the introduction of a higher field scanner. Higher fields have been demonstrated to be superior for fMRI, with prevailing opinion dictating that data cannot be meaningfully combined across different systems. Recently, multi-site studies across different systems with same field strength have demonstrated that this opinion is not correct. However, the data availability in multi-site studies is still limited since the effect of mixture data from different field strengths upon multi-site study results are not identified quantitatively or qualitatively. Considering this issue, a mixing study on fMRI auditory data from 1.5 Tesla and 3 Tesla has been conducted, investigating group analysis performance as a function of relative fraction of subjects included from each of the two different field strengths. Contrary to the general opinion, the inclusion of lower field data does not yield as much harmful effects on group analysis as the most researchers hypothesized. Thus, the mixing study contributes to not only uncovering the consequences of group analysis performance with regard to the inherent system-related signal properties, but also increasing the availability of potential fMRI data for large-scale studies. The second project is the development of a novel fMRI analysis procedure. This project is motivated by reevaluating assumptions of a currently widely-used fMRI analysis method—general linear model (GLM). GLM assumes the shape of fMRI signal, hemodynamic response function (HRF), is same across space, stimulus types and subjects, which is not a valid assumption. GLM also makes an invalid assumption of which noise is statistically independent of space and time. The second project generalizes the assumptions of GLM by incorporating the knowledge of hemodynamic responses from particular data via previously proposed probabilistic discriminant analysis (PDA). In a broad sense, PDA framework consists of iterative fMRI volume segmentation and statistical significance evaluation. The second project extends and refines the segmentation and the statistical inference of PDA. In the first component of the second project, multiple HRF models were incorporated in HRF estimation for single-subject analysis to minimize potential bias of the segmentation arising from mis-modelling of cluster-level HRFs. PDA for multi-subject data was also proposed to account for the inhomogeneity of HRFs over types of tasks and subjects, utilizing estimated HRFs in both volume segmentation and statistical significance evaluation. The algorithms of my second project were validated using synthetic data and assessed using human data, exhibiting an advantage in sensitivity and specificity in both single-subject analysis and group analysis.

Degree

Ph.D.

Advisors

Talavage, Purdue University.

Subject Area

Biomedical engineering|Electrical engineering

COinS