Research On Feature Extraction For Structural Magnetic Resonance Imaging

Alzheimer disease (AD), a progressive neurodegenerative disorder, is the most common cause of dementia in the elderly. Currently, the researchers still have not reached a consensus on the exact pathogenesis of AD. The progression of AD is gradual and it will affect areas of cognition including memory, attention, language, and problem solving. In the later stages of AD, the cognition function of AD patients will get more impaired and will die of infection or other complications at last. With the growth of the aging population in our country, the prevalence of AD is expected to get higher in the nearly future and the cost of the disease is inestimable. Not only is the financial burden substantial, but the psychologic and emotional burden on patients and their families is even greater.Clinical and neuropathological studies have greatly advanced our knowledge of the pathophysiology and progression of AD. However, a definite diagnosis of AD can only be made after a study of brain tissue at autopsy. Therefore, a facile and non-invasive diagnostic method is in urgent need. With the development of medical imaging technology, the clinical evaluation based on neuroimaging has already become a pivotal part in the clinical diagnosis of AD, where the structural magnetic resonance imaging (sMRI) is widely utilized. The sMRI scans can record the changes of biomarkers in brain structure from early memory decline to more severe clinical stages. The sMRI data refreshes our knowledge about AD and will influence the diagnosis and guide the treatment.The traditional processing and analysis methods of sMRI data are complicated, time-consuming and require the users to equip with rich experience and knowledge. A fully-automated processing and analysis method is in badly need which can detect the potential AD biomarkers in the brain structures of subjects. The signal processing methods, such as blind signal separation, multilinear subspace learning and machine learning, have been developed maturely, providing a technological possibility for the fully-automated processing and analysis method of sMRI data. With the help of these signal processing methods, vital information can be extracted from sMRI data for the diagnosis of AD.Independent component analysis (ICA), a blind source separation method, becomes more and more popular in analyzing the sMRI data of AD. In these studies, with the assumption that the sMRI images of subjects are independent, features are extracted from these images by ICA for further classification and diagnosis. However, in this feature extraction model, a set of sMRI images are needed. In other words, this ICA feature extraction scheme can not extract features from only one subject’s sMRI data and therefore can not diagnose for only one subject. This shortcoming makes this diagnosis scheme not satisfied the demand of clinical diagnosis, which requires that the diagnosis method can be product on every new subject immediately. To solve this problem, this paper proposed a new ICA feature extraction scheme which can extracts features from every single subject’s sMRI data for further diagnosis. In this ICA feature extraction scheme, we assume that the voxels of each sMRI image are independent. A set of training sMRI data are used to train a separating matrix and a classification model, which can be used for the feature extraction and classification of every new sMRI data. This diagnosis scheme can be conducted on every subject immediately and satisfied the demands of clinical diagnosis. The simulation results show that the performance of new model is as good as the traditional one, but more suitable for the clinical use.Some two-dimensional feature extraction methods, including ICA, have shown good performance on analyzing sMRI data. However, all these two-dimensional methods vectorize the origin three-way sMRI data, leading to high dimensional vectors which may cause under sample problem and a great loss of the structural information of the data. To overcome these disadvantages, this paper presents a classification scheme that combines the uncorrelated multilinear principal component analysis (UMPCA) and Laplacian score (LS) methods, which are kwon to be effective to the structural correlation preserving and redundancy reduction of the AD-related features hidden in the sMRI data. UMPCA extracts features directly from the tensorial sMRI data, preserving most of their structure information. Besides, the extracted information will be less redundant after feature selection procedure by the LS algorithm which removes irrelevant components and reduces computational cost. The simulation results show that the classification accuracy of the proposed UMPCA-LS method is higher than other current classification methods.