Title: Machine Learning Methods for Magnetic Resonance Imaging Analysis
Co-Chairs: Professor Tailen Hsing, Assistant Professor Xuanlong Nguyen
Cognate Member: Professor Douglas C. Noll
Members: Professor Kerby Shedden, Professor Naisyin Wang, Assistant Research Scientist Scott Peltier
Abstract: The study of the brain and its connection to human activities has been of interest to scientists for centuries. However, it is only in recent years that medical imaging methods have been developed to allow a visualization of the brain. Magnetic Resonance Imaging (MRI) is such a technique that provides a noninvasive way to view the structure of the brain. Functional MRI (fMRI) is a special type of MRI, measuring the neural activity in human brain. The aim of this dissertation is to apply machine learning methods to functional and anatomical MRI data to study the connection between brain regions and their functions.
The dissertation is divided into two parts. The first part is devoted to the analysis of fMRI. A standard fMRI study produces massive amount of noisy data with strong spatio-temporal correlation. Existing methods include a model- based approach which assumes spatio-temporal independence and a data-driven method which fails to exploit the experimental design. In this work we propose a Gaussian process model to incorporate the temporal correlation through a model-based approach. We validate the method on simulated data and compare the results to other methods through real data analysis. The second part covers the analysis of anatomical MRI. Anatomical MRI provides a detailed map of brain structure, especially useful for detecting small anatomical changes as a result of disease process. The goal of anatomical MRI analysis is to train an automated classifier that can identify the patients from healthy controls. We propose a multiple kernel learning classifier which will build classifiers in small regions in the segregating step and then group them in the integrating step. We study the performance of the new method using simulated data and demonstrate the power of our classifier on disease-related data.