Statistical Methods for Analysis of Structural Magnetic Resonance Imaging in Patients with Multiple Sclerosis

Multiple sclerosis (MS) is an inflammatory disease of the brain and spinal cord characterized by demyelinating lesions. Structural magnetic resonance imaging (sMRI) is a medical imaging technique that is sensitive to these lesions. Quantitive analyses of MRI, such as the number and volume of MS lesions, are essential for diagnosing the disease and monitoring its progression. In addition, the formation of these lesions, a complex process involving inflammation, tissue damage, and repair, is also important for diagnosing and monitoring the disease. While sMRI is sensitive to lesion activity, there is surprisingly poor association between clinical findings and the radiological extent of involvement on MRI using traditional volumetric measures. This phenomenon is referred to as the clinico-radiological paradox. The work in this thesis is an effort to bridge this clinico-radiological paradox and link the longitudinal findings on structural MRI in patients with MS to disease-modifying treatment and other clinical information. Chapter 2 of the thesis is an introduction to sMRI data. Chapter 3 and 4 of the thesis deal with MS lesion segmentation using multi-sequence structural MRI. Chapter 5 is a culmination of this work. The lesion segmentation technique explored in Chapter 3 and 4 is extended to build a pipeline to extract longitudinal intensity information, or lesion profiles, from lesions in multi-sequence sMRI. A PCA regression model is then introduced to relate the longitudinal lesion profiles to disease-modifying treatment and other clinical information in an attempt to link the information from sMRI to clinical information. In addressing these clinical issue, this thesis also contains a number of biostatistical contributions: the design and analysis of expert rater trials, data reduction techniques for high dimensional and longitudinal data through principal component analysis (PCA) regression models, and the comparison of supervised learning algorithms