Feature Learning Method Research Based On Structural Magnetic Resonance Imaging For Classification Of Brain Diseases

Magnetic resonance imaging(MRI)technology makes it possible to measure the morphology and functional information of the human brain.We can use it to analyze brain abnormity for revealing the physiological and pathological mechanisms of the brain diseases.It has a very important significance in clinic diagnosis and treatment of various brain diseases.Alzheimer’s disease is a progressive,disabling neurodegenerative disorder by the presence of brain atrophy and neurofibrillary tangles.The disease is one of the most fatal threats to people in the 21 st century,and affects the health and life quality of patients seriously.Since the human brain is a complex system,it is still an open question that how to characterize the brain reasonably and comprehensively.Recently,machine learning technology,which has been successfully applied to the prediction and diagnosis of brain disease,is helpful for us to understand the brain mechanism.Based on structural magnetic resonance imaging data we examine feature learning methods for mine brain discriminative information.The main research works are as following:1.We propose a manifold regularization sparse feature selection method based on label information.The label information is added in the Euclidean distance formula to form a novel manifold regularization term.It can maintain the compactness of intra-class samples and the separability of inter-class samples for selecting more discriminative features.We carry out three binary classification tasks and one multi-class classification task on Alzheimer’s Disease Neuroimaging Initiative database.Experimental results demonstrate the effectiveness of the proposed method.The discriminative features can serve as useful biomarkers for the diagnosis of Alzheimer’s disease and mild cognitive impairment.2.We propose a brain network connection method based on k-nearest neighbors using multiple structural features of sMRI.Since the brain is a complex interaction system,the traditional network connection is generally quantified by the structural difference pattern between pairs of brain regions.In this way,the interaction between multiple brain regions is ignored leading to the rough quantification of brain regions connection.So we use the brain region’s k-nearest neighbors to enrich the brain region information and propose a brain network connection method based on k-nearest neighbors.A new brain vector is reconstructed by using adjacent brain regions and represented the original brain region.The connection relationship between pairs of two original brain regions is represented by the connection relationship of two new brain’s vectors,and thus this relationship is more accurate.The experiments show our method can achieve good performance,as well as the significant difference connections and hub brain regions are consistent with the physiological and pathological mechanisms of Alzheimer’s disease.So,it has important biomedical significance for the diagnosis of mild cognitive impairment.3.We propose a weighted sparse brain network connection method using multiple structural features of sMRI.In previous studies,the method of constructing individual network by sparse representation has certain limitation: sparse representation does not guarantee to be local which lead to lose the important information;the target brain region may be represented by more different brain regions.Given this,we integrate the sparsity and locality of data,and propose a weighted sparse brain network connection method.The weighted sparse coefficient is used to quantify the interacting relationship between multiple brain regions,which can better map the "one-to-many" interaction of the brain regions.The experimental results show that our method has higher classification performance than the sparse representation method.For mild cognitive impairment classification,The proposed classification accuracy is improved by about 20%.Moreover,we find disease-related hub brain regions and differential relations.It is of great significance to clinical diagnosis.