Print Email Facebook Twitter Dimension reduction methods for classification; MRI-based automatic classification of Alzheimer's disease Title Dimension reduction methods for classification; MRI-based automatic classification of Alzheimer's disease Author Van Giessen, A. Contributor Vandal, A.C. (mentor) Jongbloed, G. (mentor) Loog, M. (mentor) Faculty Electrical Engineering, Mathematics and Computer Science Department Delft Institute of Applied Mathematics Date 2012-03-26 Abstract Alzheimer's disease (AD) is a type of dementia which is difficult to diagnose based on clinical observations. Many automated classification algorithms are being developed to aid in the diagnosis. In such algorithms, principal components analysis (PCA) is a popular tool to reduce the dimension of data, get rid of noise and redundancy and thereby improve the classification. As PCA is a form of unsupervised learning, i.e. it relies entirely on the input data itself without reference to the corresponding target data, it does not make use of any available information about group structure. Applying PCA to data containing high within-group inter-subject variability, and possibly only subtle differences between the groups, as is the case for people with stable mild cognitive impairment (MCI) and people with MCI which will convert to being diagnosed with AD, might not improve classification results much since the outcome of PCA will be spoiled by the high variance between the subjects. In this study new methods will be introduced that take into account the available information on group structure to select features or principal components. One approach is based on minimizing the similarity between the principal components of two groups using the concept of computing angles between subspaces generated by these principal components, while the other is based on logistic regression. These novel methods are evaluated and compared to the conventional methods by their classification performance on data consisting of brain volumes, which were extracted from MRI data obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI), using logistic regression classification, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and support vector machines (SVMs). Given only 10-15% of subjects with MCI convert to AD each year, it is necessary to correct for imbalance in the data. This is done for logistic regression by optimizing the threshold and in SVM by optimizing a cost parameter that assigns different costs to each class. Compared to using all features as well as to the conventional application of PCA, where the dimension is reduced by selecting the first principal components accounting for a certain percentage of the variance in the data, application of several of these supervised dimension reduction methods to the ADNI data shows improved classification results. To reference this document use: http://resolver.tudelft.nl/uuid:89bf4948-87b4-4a9a-853b-8b828ae7ed24 Part of collection Student theses Document type master thesis Rights (c)2012 Van Giessen, A. Files PDF Thesis_AM_250312.pdf 985.58 KB Close viewer /islandora/object/uuid:89bf4948-87b4-4a9a-853b-8b828ae7ed24/datastream/OBJ/view