Probabilistic discriminant analysis for functional MRI and model-based tuning algorithm for MRI coil array

Liang Liu, Purdue University

Abstract

This Ph.D. thesis includes two separated parts: (1) Probabilistic Discriminant Analysis for fMRI analysis. A novel functional Magnetic Resonance Imaging (fMRI) data analysis framework is developed to detect neural activated brain regions based on an iterative segmentation that exploits probabilistic discriminant analysis (PDA). Given hemodynamic response function (HRF) model and experiment paradigm, the PDA framework alternates between estimation of the hemodynamic response (HDR) signals and noise component for clusters of brain voxels and refinement of the voxel segmentation. The latter process is effected through minimization of misclassification risk. The ensemble of HDRs helps to reduce the multiple comparison problems that exist in conventional univariate analysis of fMRI data. The statistical parametric mapping is generated by a general linear model (GLM) approach, but operating on the final optimized clusters rather than individual voxels. Assessment of the PDA framework using synthetic activation data embedded in temporally and spatially correlated noise shows that when optimized using a linear discrimination function, the PDA method outperforms the traditional regular GLM and independent component analysis (ICA) approaches. Comparison among these methods of results from human data suggests that the PDA framework can enhance both sensitivity and specificity, yielding more reliable detection of activated brain regions for fMRI studies. (2) Model-based Tuning Algorithm for MRI coil array. A model-based tuning method for MRI receiver coils is developed that optimal tunable component values can be predicted and manual operations in coil tuning process can be minimized. For multi-element/channel MRI receiver arrays, it is generally difficult to achieve effective isolation among all channels by only overlapping neighborhood coil loops. Applying low input impedance preamplifier after matching network is the major technique used for non-adjacent coil loop isolation, also known as preamp decoupling method. For MRI coil array tuning, besides matching loaded coil impedance for optimized the preamplifier noise figure at resonant frequency, maximizing the preamplifier decoupling impedance is also essential to constrain current flow in the coil loop thus reduce inductive coupling between coil channels. Both matching and preamplifier decoupling impedances can be adjusted by tunable components (typically capacitors) in the coil circuit. Given that circuit component values vary from nominal values, tuning operations beyond initial coil assembling are necessary so that coil elements can meet tuning specifications. Conventional manual tuning approach requires human experience and iterates over alternate steps of tuning the matching impedance and preamplifier decoupling impedance seperately. It costs extra tuning operations and capacitors in order to meet both impedance specifications simultaneously. The proposed model-based tuning method utilizes a parameterized coil model to estimate component values from impedance measurements and predict values of tunable components such that all specifications can be satisfied. From both simualtion and experiment on coil array tuning, the proposed method takes less tuning operations and also achieves more optimal matching performance than manual tuning method.

Degree

Ph.D.

Advisors

Talavage, Purdue University.

Subject Area

Electrical engineering

COinS