Novel approaches based on decomposition methods for detecting MRI patterns in the progression of Alzheimer's Disease

Abstract

Image-based computer aided diagnosis (CAD) systems have significant potential for screening and early detection of brain diseases. In this sense, this PhD work is motivated by the development and the implementation of novel CAD systems based on several pattern recognition/classification techniques for the early detection of neurodegenerative diseases. In particular, the dissertation is focused on the analysis of the most relevant one, the Alzheimer’s disease (AD), by the use of structural magnetic resonance imaging (sMRI) techniques. The proposed CAD systems are based on several processing steps including segmentation of brain tissues (gray matter and white matter tissues), feature selection techniques such as t-test model, feature extraction techniques such as; Partial least squares (PLS), Principal Component Analysis (PCA), Independent component analysis (ICA) and Non-Negative Matrix Factorization (NNMF) techniques, and an automatic classification technique, such as the support vector machine (SVM). Most of these thesis contributions are included within the feature extraction techniques and its application to the development of automatic CAD systems for early detection of AD. The first proposed CAD system is based on the PLS approach that extracts the relevant features to characterize the AD pattern. This technique decomposes two sets of variables into the product of two matrices called scores and loadings according to a criterion of covariance maximization. In this work, these variables are those formed by the structural magnetic resonance images under study and the labels of these images. After the decomposition of these sets, the scores are used as feature vectors for the classification step. The second proposed CAD system for AD detection is based on the PCA approach, as a feature extraction technique for sMRI brain images. This approach reduces the original high-dimensional space of the brain images to a lower dimensional subspace. PCA generates a set of orthonormal basis vectors, known as Principal components (PCs), that maximizes the scatter of all the projected samples, which is equivalent to diagonalize the covariance matrix through the eigenvalue of these components. The third CAD system is based on the Independent component analysis (ICA) approach. At the beginning, a “template” image is computed as the average of healthy subject images or as the difference between normal and pathological images. Then, the ICA algorithm is applied to extract the maximally spatially independent components (ICs) revealing patterns of variation that occur in the dataset under study. The last method proposed is based on the NNMF approach, as a useful decomposition technique of multivariate data that solves the problem of finding non-negative matrices. These feature extraction techniques successfully solved the small sample size problem by obtaining only the relevant information related to AD. This process is known as a dimensionality reduction and it improves the prediction accuracy of CAD systems, specifically, in the early stage of AD. In this way, considering the fact that an early diagnosis of AD is crucial, classification experiments were performed not only to distinguish Normal Control (NC) and AD subjects but also to differentiate NC from a transitional phase between being cognitively normal and having an AD diagnosis. This later phase is called Mild Cognitive Impairment (MCI). The proposed feature extraction methods have been combined with SVM classifiers and the accuracy rates of the resulting CAD systems have been estimated by means of a sMRI database from the Alzheimer disease neuroimaging initiative (ADNI). Furthermore, a kfold-cross validation technique was applied to these systems in order to tune the classifier parameters and to estimate its performance. The obtained results demonstrate the effectiveness and the robustness of the proposed CAD systems (with accuracy value around 90%) compared to previous approaches such as the one based on Voxel-As-Features (VAF) technique.