Research On Representation Learning Of Brain Network And Its Application

Machine learning techniques have been successfully applied in various fields,including image processing and medical image analysis.Recently,researchers have also applied machine learning to the analysis(classification)of brain networks based on functional magnetic resonance imaging(fMRI).However,how to abstractly and accurately express brain network is still an important problem and also the basis of subsequent analysis in brain network because of essentially high-dimensional structured data.Based on this background,this paper carried out the research work of representation learning with functional brain network,which includes the following three parts:(1)We proposed a novel adaptive thresholding(called WDT)method for functional brain networks.Threshold processing is a very basic brain step for network analysis.The traditional methods generally adopt a single value or a percentage of edges to threshold the whole brain networks,ignoring the diversity of connection between brain regions,i.e.,difference of strength(i.e.,weight)of connections between different brain regions.So we should adopt different thresholds for different connections.In addition,there is no golden standard to determine the optimal threshold or percentage.In practice,researchers usually need to try a large number of possible values to determine the optimal threshold.Different with existing methods,the proposed WDT method can make full use of the distribution information of weight of connections in different brain regions,and determine an optimal threshold for each connection adaptively,so as to describe the brain networks more accurately.Specifically,we first split the training samples into two groups according to their class labels(i.e.,patient and normal control).Then,for each connection,the corresponding threshold is automatically determined using the weight distribution of weight of that connection across different sample groups.Thus,the proposed WDT method can determine different thresholds for different connections in different brain regions,preserving the diversity information of connections.(2)We proposed a representation learning method of integrating multiple properties of brain network.Considering that a single feature of brain network characterizes a single kind of property of brain network,and different network features describe different network characteristics,which may contain useful and complementary information to further improve the performance of brain network analysis.Therefore,following the previous studies,we extract different network measures,which characterized connectivity properties of brain network from different sights,from the thresholded brain network using the WDT method.Furthermore,we propose a representation learning method of fusion of multiple network features using multi-kernel support vector machine(multi kernel-SVM)technology.(3)The proposed model is applied to the brain diseases analysis.The performance of the proposed model was evaluated on two public available fMRI-based brain disease data sets,i.e.,the Alzheimer's Disease Neuroimaging Initiative(ADNI)and Attention Deficit Hyperactivity Disorder(adhd-200).Experimental results show that the proposed method can further improve the performance of brain disease classification.