Brain Network Classification Method With Multiple Features Fusion Based On Minimum Spanning Tree Analysis

Brain network,as a tool widely used to identify abnormal topology properties,is one of the research focuses in the field of brain science.However,Conventional network analysis to compare groups are hampered by differences in network size,density and suffer from normalization problems.Therefore,the minimum spanning tree(MST)is a new important tool for the study of brain network,which is widely applied in the research of neuropsychiatric disorders.The minimum spanning tree is a subgraph of the initial network,mathematically defined as a subnet that connects all nodes,and the sum of all connections is minimal and not looped.The minimum spanning tree analysis method avoids the deviation of the methodology and solves the shortcomings by using conventional graph theory method.In the process of constructing the resting state functional brain network of depressive disorders and normal subjects,this study used the MST method.Then we analyzed the minimum spanning tree from brain regions and subgraphs to explore the difference between the depression patients and the normal controls.We also combined the brain region features and subgraph features andselected multi-kernel SVM as a classifier.The innovative work in this paper goes as follows:First,the minimal spanning tree analysis method is introduced into the brain functional network construction.The minimum spanning tree method is a new method to construct brain network,which has been widely used in the related research of neuropsychiatric diseases in recent years.This study constructed the functional brain network by minimum spanning tree method,which avoids the problem of setting the threshold of traditional graph theory.This paper based on resting fMRI data to build the minimum spanning tree,calculate and analyze the minimum spanning tree network topology.We found that the depression’s MSTs were more tend to random networks by comparing the global attributes with normal controls.Compared with normal controls,MDD patients appeared some significant abnormal brain regions,that were belong to Limbic-Cortical-Striatal-Pallidal-Thalamic loop,which is considered to be the major pathological circuit of depression.The results are consistent with the experimental results of the traditional graph theory,indicating that the analysis of depression using the minimum spanning tree method is reliable and feasible.This study of new methods of depression can not only enable us to understand the pathogenesis of depression,but also provide some help for clinical intervention and treatment.Secondly,the study of subgraph features which extracted from minimum spanning tree networks makes the description of the brain network not limitedto a single brain region.Traditional classification methods just pick up a series of features from the brain network and then connect these extracted features into a feature vector,which can be used for subsequent classification.However,an obvious drawback of this approach is that some of the useful network topology information(including the topology structure of the sample and the common topological structure between samples)may be lost,and then affect the performance of the classifier.The research methods based on subgraph does not need to be limited to a single brain region in the description of the brain network,which not only retain the topological information of the original sample,but also does not lose the original discriminant information.In this experiment,the subgraph features of the minimum spanning tree network were studied.The results showed that the abnormal subgraph patterns and connections in the MST networks of depressive patients were consistent with the existing findings.This finding is of great significance to medical-assisted diagnosis.Finally,a novel brain network classification method based on minimum spanning tree network for multiple features fusion is proposed.The brain network is a complex structure.Thus the single feature can not be fully acquired its biological characteristics.Different types of characteristics are used to measure the change of brain network.On the one hand,brain region features are sensitive to the change of single brain region.On the other hand,the topological information of the whole brain network can be characterized by subgraphfeatures.Here,we computed degree,eccentricity,and betweenness centrality of MST networks and selected the brain regions with significant between-group difference as local network features.And the most discriminant subgraphs were extracted from MST networks as subgraph features.Then,we combined the two types of features,and chose a multi-kernel support vector machine(SVM)as the classifier to achieve better classification performance.In this paper,the brain regions of the three local indicators with significant between-group difference were selected as the brain features,then the brain features and subgraphs were combined and multi-kernel Support Vector Machine was selected as our classifier.The current study demonstrated that this novel classification method could effectively improve the classification accuracy and had better interpretability.The results showed that brain network can make use of different types of features to complement each other.The new method of this paper has a certain reference value for the diagnosis of clinical diseases,which can help to diagnose the depression and improve the accuracy and efficiency of diagnosis.