| The human brain is a complicated system made up of a great quantity of neurons,neurons clusters,and brain regions.The brain network was modelled by the combination of brain imaging technology and complex network theory,which provided a new perspective for exploring the working mechanism and pathological mechanism of the human brain.As we all know,there is multivariate and complicated high-order information in the nervous system.On this basis,researchers introduced hyper-network theory to create a brain functional hypernetwork model,thus describing the complex high-level information interaction in the human brain.In recent years,this model has been widely used in simulating multivariate interactive information in the human brain and in diagnosing mental illnesses of the brain.However,due to the limitation of the method,the existing hyper-network model could not express the inherent priori group structure problem in the human brain,and at the same time,the constructed model was not analyzed and further applied effectively.Considering these problems,based on functional magnetic resonance imaging data,we proposed corresponding solutions for the construction,analysis and application of resting-state brain function hypernetwork,so as to enrich the theoretical framework and the application system of resting state brain functional hyper-network.Finally,we achieved computer-assisted platform for the construction and analysis of the resting state brain functional hyper-network.The main content of the dissertation included the construction technology for resting state brain functional hypernetwork based on group structure,the robustness technology of resting state brain functional hyper-network,the construction technology of resting state brain functional hyper-network classification model based on multi-feature extraction and multi-feature fusion,the construction technology for high-order resting state brain functional hyper-network and the development of a computerized platform based on resting state brain functional hyper-network.Then we adopted major depression disorder as the main disease model to verify the effectiveness and feasibility of the above methods and to explore the biomarkers for early diagnosis.In addition,based on hyper-network features,we used machine learning methods to construct computer classification models,thus assisting clinical diagnosis.The main innovations of this dissertation include:(1)Propose the construction and analysis technology of resting brain functional hypernetwork based on the group LASSO and sparse group LASSO methodAiming at the problem that the existing hyper-networks cannot express the inherent prior group structure,we proposed the construction and analysis of resting brain functional hypernetwork based on the group LASSO and the sparse group LASSO method to improve the construction of hypernetwork.Both methods are self-defined group selection methods,that is,the hyper-network was constructed on the basis of the prior group structural information of the human brain.The difference is that the group LASSO method is a group-level selection method that only performs group-level selection of brain regions.The sparse group LASSO method is a bi-level group selection method in which brain regions are selected both intergroup and intragroup.We firstly used clustering method to distinguish the strongly related brains into a group,and then we adopted the traditional LASSO,group LASSO and sparse group LASSO methods to construct a functional hyper-network.Next,we applied the hypernetwork cluster coefficients to quantify the hypernetwork and non-parametric test method was adopted to select the discriminative features between groups.Finally,we employed support vector machine to build classification model.In addition,we introduced the hyper-network robustness theory to estimate the impact of robustness among different hyper-networks and utilized the Relief F method to assess the features weights obtained by different methods.Accordingly,the topology of the brain hyper-network,the robustness of the brain hyper-network,different brain areas,the classification performance and the classification weights were compared from several perspectives.The results show that the hyper-network topology constructed by the sparse group LASSO method was moderate;the group LASSO method,most lenient;and the LASSO method,strict.And the sparse group LASSO method achieved better classification performance,which indicated that effective brain spacial multivariate interactions can be quantified if the group structure existed and was properly extended.(2)Propose multi-feature extraction and multi-feature fusion classification technology based on the brain functional hyper-network modelIn light of the problem of one-sided and flat expression of single type of topological property in traditional brain functional hyper-network model,we proposed the multi-feature extraction and multi-feature fusion classification based on the brain functional hyper-network model,so that evaluating the local topology of the brain functional hyper-network in a multidimensional perspective.Fistly,a multi-feature extraction was performed,where from the point of node attributes,eleven local attributes were introduced to quantify the topology of the brain functional hyper-network constructed by the sparse group LASSO method.Secondly,the nonparametric test method was used to select the difference features of each group of topological properties,and classification was performed,and further the minimum redundancy and maximum correlation(m RMR)algorithm was adopted to evaluate the effectiveness of the features.Then,each single topological property was compared and analyzed from the perspective of classification results and feature validity.Finally,based on the topological properties that contain more classification information,the classification model was constructed using tandem fusion and multi-kernel learning fusion based on the alignment maximization algorithm.The experimental results show that,from the perspective of single topological property,three clustering coefficients defined on single nodes,two clustering coefficients defined on pairs of nodes and the average shortest path achieved better classification performance.From the perspective of the fusion features,the fusion features composed of the above topological properties were superior to the fusion feature composed of all topological properties and any single properties.Moreover,the fusion feature constructed using the multikernel learning method achieved higher classification accuracy.These results suggeted that fusion features could contain information of a variety of different topological attributes and avoid the one-sided problem.Compared with the tandem fusion method,the multi-kernel learning method could further enhance the differences between two groups of subjects,obtain more effective classification model,and improve the disease diagnosis of the hyper-network.(3)Propose the construction and analysis technology of high-order resting brain function hyper-networkAiming at the dynamic changes of brain connections in the resting state,we proposed highorder resting state brain functional hyper-network construction to simultaneously characterize the dynamics and diversity of working mechanisms of the human brain.Specifically,we used the sliding time window method to obtain the relevant time series.Based on the relevant time series,the sparse group LASSO method was used to construct a high-order brain hyper-network,thus reflecting the time-varying properties of the hyper-network.Then,we introduced the hyper-edges as dynamic sub-graph features to represent the global information of the high-order hyper-network,and applied the frequent scoring feature selection method to select the dynamic discriminant sub-graph patterns.Next we adopted the Weisfeiler-Lehman subtree to quantify the discriminant subgraph,and further performed classification.Meanwhile,we extrcated the local topological properties that contain more information on the classification of brain diseases(obtained by in the previous chapter)as local property features,and used non-parametric permutation tests to select discriminative features to perform the classification.Finally,the multi-kernel learning method based on the alignment maximization algorithm was applied to fuse discriminative sub-graph features and local attributes in order to construct classification model.In addition,the Relief F algorithm was performed to evaluate the classification weight.Furthermore,the reliability of the proposed method was verified from four aspects of abnormal brain areas,dynamic subgraph patterns,classification performance and feature weights.The results show that the high-order hyper-network achieved the better classification performance,which indicated that if we want to more accurately simulate the complex working mechanism of the human brain and accurately identify the biomarkers of mental illness,not only the spatial and multiple interaction relationships of the human brain need to be considered,but also the time-varying properties of the human brain need to be taken into account.(4)Achieved the design and development of brain functional hyper-network construction and analysis platformIn view of the coding difficulties in diagnosing brain disease using hyper-network theory for non-computer researchers,we achieved the development of brain functional hyper-network construction and analysis platform(Brain Hypernet Construction).Specifically,we developed the Brain Hypernet Construction platform by integrating the construction analysis techniques in this dissertation and combining the traditional and other existing hyper-network research of our team.The graphical user-friendly interface was used to help researchers to perform tasks such as constructing,analyzing,and employing hyper-network in a quick,easy,and flexible manner,and to improve the repeatability of diagnosing brain disease based on hyper-network theory. |
