Multi-Class Pattern Analysis On Human Brain MRI Dataset

In recent years, pattern recognition methods have been widely applied to the analysis and study of the Magnetic Resonance Imaging (MRI) data. With the deep study of the MRI data, multi-class problem for MRI data gradually came forth, which puts forward higher request for the study of work mechanism and disease pathology of the brain and also provides application for multi-class pattern recognition analysis. In this study, we applied pattern recognition methods to multi-class problem and analysis of the MRI data, expecting to obtain meaningful analysis results.In this paper, three groups of resting-state functional MRI (fMRI) data were acquired from schizophrenia patients, their healthy siblings and healthy controls, respectively. Principal component analysis (PCA) combined with nonlinear support vector machine (SVM) was applied to functional connections of the three groups. One-against-one classifications were employed among the three groups. The classification results have proven the inheritance of schizophrenia from the view of pattern classification.Besides, in order to obtain the brain regions which contribute to the inheritance of schizophrenia, we analyzed the above resting-state fMRI of the three groups using group ICA, resulting in 20 independent components. Using all components as feature congregation and each component as feature sub-congregation, we applied feature fusion, PCA and linear SVM to one-against-one classifications among the three groups, and obtained the consistent results with the above study, that is, proving the inheritance of schizophrenia. While, the most important is that we found the crucial brain regions which contributed to the inheritance of schizophrenia.At last, we analyzed three groups of structural MRI (sMRI) data acquired from Alzheimer's disease patients (AD), mild cognitive impairment patients (MCI) and healthy controls. Improved recursive feature elimination method based on Support Vector Machine (SVM-RFE) was first employed in the sMRI data for feature extraction, then relevance vector machine (RVM) was applied to the extracted features for multi-class pattern classifications. From the results of two-class pattern classifications, we found that the obtained features through pattern recognition were consistent with the results of statistical difference study, which mainly focused on the hippocampus area and its around areas, and these areas take charge the memory function of human beings. Besides, we also did some exploring study on multi-class pattern classification for the sMRI data, including"one-against-one"multi-class classification and"one-against-rest"multi-class classification. Through these multi-class classifications, we came to the conclusion that the MCI group already showed highly cognitive impediment and were converting to the Alzheimer's Disease. Through multi-class pattern recognition on the MRI dataset, we not only have deep comprehension on the multi-class problem of the MRI data, but also obtain the physiological and pathological results of the MRI dataset from the view of pattern recognition, which made good contribution to the application of pattern cognitive methods to brain research.