A Research Of Pattern Recognition Methodology For Brain Diseases & Disorders Classification With Magnetic Resonance Imaging

Magnetic resonance imaging(MRI) has been widely applied in mapping the brain tissue structure and functional activities, however, it is still difficult to answer that how can we determine efficient diagnostic imaging and anlysis methods used for computer-aided diagnosis utilizing many kinds of methods and application of MRI, thus a clinically feasible diagnosis scheme worths to be studied sysmetically. The research motivation of this dissertation was to evaluate and validate the practical performace of classifiers combined with different feature selection and extraction methods for different MRI modalities, and multi-centre data, then, a novel algorithm was proposed based on the experience. The emphasis was on showing practical performance and new optimised design clinical pattern recognition methods for MRI from different viewpoints.The main work was divided into four procedures: The first one was to transfer the research discovery on single centre data of Schizophrenia to the data anlysis on multiple centres. Once it has been successful, robust methodology resulted from multi-centre validation was achieved, then the data of different diseases and MRI modalities from one MRI scanner was used to validate basic pattern recognition methods and explore the necessity of feature extraction procedure, in which there were diseases including Parkinson’s Disease(PD), Schizophrenia(SZ), Depression,Bipolar Disease(BP), and the modalities including 3D T1 structural MRI and BOLD-functional MRI. The second step was to try some advanced feature selection methods to promote the classifier accuracy as well as the features mapping without compromise on lossing the classification accuracy and afterwards a novel feature selection method was proposed based on sparse model,structural and randomized approach integrated with some useful frameworks already validated in this dissertation. The third step was applying the new algorithm and a connectivity-based method to the Parkinson’s Disease(PD),of which the molecular pathology and pathogenesis are more determined, for calculating discrimative voxels between patients and controls as parallel analysis and discussions.What is reported in the first time is that: 1.The result and performance in several centres performs consistently, for instance,the new performance with modified methods seem to be improved synchronously; 2.Different modalities contributed differently to the diagnosis of different diseas,the best imaging solution for each disease varies; 3.New algorithm-“Randomized Structural Sparsity” was proposed. 3.When used in Parkinson’s Disease, the explanable result of new algorithm has been very consistent with the expected brain regions according to the physiological and pathological mechanisms of PD.