Abstract

Researchers in the field of functional neuroimaging have faced a long standing problem in pre-processing low spatial resolution data without losing meaningful details within. Commonly, the brain function is recorded by a technique known as echo-planar imaging that represents the measure of blood flow (BOLD signal) through a particular location in the brain as an array of intensity values changing over time. This approach to record a movie of blood flow in the brain is known as fMRI. The neural activity is then studied from the temporal correlation patterns existing within the fMRI time series. However, the resulting images are noisy and contain low spatial detail, thus making it imperative to pre-process them appropriately to derive meaningful activation patterns. Two of the several standard preprocessing steps employed just before the analysis stage are denoising and normalization. Fundamentally, it is difficult to perfectly remove noise from an image without making assumptions about signal and noise distributions. A convenient and commonly used alternative is to smooth the image with a Gaussian filter, but this method suffers from various obvious drawbacks, primarily loss of spatial detail. A greater challenge arises when we attempt to derive average activation patterns from fMRI images acquired from a group of individuals. The brain of one individual differs from others in a structural sense as well as in a functional sense. Commonly, the inter-individual differences in anatomical structures are compensated for by co-registering each subject's data to a common normalization space, known as spatial normalization. However, there are no existing methods to compensate for the differences in functional organization of the brain. This work presents first steps towards data-driven robust algorithms for fMRI image denoising and multi-subject image normalization by utilizing inherent information within fMRI data. In addition, a new validation approach based on spatial shape of the activation regions is presented to quantify the effects of preprocessing and also as a tool to record the differences in activation patterns between individual subjects or within two groups such as healthy controls and patients with mental illness. Qualititative and quantitative results of the proposed framework compare favorably against existing and widely used model-driven approaches such as Gaussian smoothing and structure-based spatial normalization. This work is intended to provide neuroscience researchers tools to derive more meaningful activation patterns to accurately identify imaging biomarkers for various neurodevelopmental diseases and also maximize the specificity of a diagnosis.

Publication Date

4-5-2013

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Stefi A. Baum

Advisor/Committee Member

Vince D. Calhoun

Advisor/Committee Member

Nathan D. Cahill

Campus

RIT – Main Campus

Share

COinS