Dimensionality Reduction for fMRI Diagnostic Systems

  • Reducing fMRI Dimensionality

  • Author / Creator
    Sidhu, Gagan
  • Functional Magnetic Resonance Imaging (fMRI) measures the dynamic activity of each voxel of a brain. This dissertation addresses the challenge of learning a diagnostic classifier that uses a subject’s fMRI data to distinguish subjects with neuropsychiatric disorders from healthy controls. fMRI intrinsically possess spatial and temporal dimensions, given by a waveform over hundreds of time points at each of 10^5 spatial locations. Given training data of only dozens to hundreds of subjects, standard learning algorithms will over-fit – i.e., do well on the training data, but poorly on novel instances. We address this by reducing the dimensionality, using several variants of Principal Component Analysis (PCA). We evaluate the per- formance of the PCA Variants on two datasets: Attention-Deficit Hyperactivity Disorder (ADHD) [a large public dataset of 668 subjects, used for the ADHD200 competition] and First Episode Psy- chosis [involving 34 subjects]. Our empirical studies show that using non-linear PCA to reduce fMRI dimensionality over both the spatial and temporal dimensions is statistically better, with respect to the classification task, than using a linear mapping to reduce over only the spatial or only the temporal dimension.

  • Subjects / Keywords
  • Graduation date
    Fall 2012
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.