Computer-aided diagnosis of alzheimer’s disease and mild cognitive impairment with MARS/CMARS classification using structural MR images

Çevik, Alper
Early detection of Alzheimer’s disease (AD) and its prodromal stage, amnestic mild cognitive impairment (MCI), has drawn remarkable attention in recent years. Despite the impressive developments in fields of image analysis, pattern classification, and machine learning, no computer-aided diagnosis system has yet been a part of the clinical routine to diagnose the AD. This thesis study aims to propose a thorough procedure which involves detecting the early signs of disease-originated deformations by fully-automated analysis of structural brain magnetic resonance images (MRI). A comprehensive review including the taxonomy of related biomarkers and state-of-the-art techniques is introduced. Proposed methodology involves extraction of voxel intensity-based features (such as tissue probability maps) through segmenation and registration of brain MRI volumes. Voxel-based morphometry framework is employed to provide one-to-one correspondance between the images. Quality of the feature set is evaluated by an analysis including other approaches such as feature-based morphometry. A novel hybrid procedure involving both statistical analysis and utilization of domain knowledge is proposed for feature selection. Performance of the method is compared with these of well-known dimensionality reduction techniques. Multivariate adaptive regression splines (MARS) and Conic MARS (CMARS) were utilized for construction of the class-separating hyperplanes through a parameter optimization procedure involving cross-validation. This study is the first-time engagement of both MARS and CMARS algorithms in field of medical image analysis. Qualitative and quantitative evaluations of classifier performances were presented including a comparison with benchmark studies in the field. Promising results are acquired through the tests performed on Alzheimer’s Disease Neuroimaging Initiative (ADNI) data.