Research On Feature Learning And Classification Of Brain Functional Networks In Mild Cognitive Impairment

Brain functional network（BFN）is a very important research content in cognitive neuroscience.Nowadays,it has become a tool for humans to understand and study the brain,and it is widely used in the study of the pathogenic mechanism and auxiliary diagnosis of brain diseases.In this paper we use normal subjects and mild cognitive impairment（MCI）subjects as the research objects,and use the resting brain functional magnetic resonance imaging（f MRI）data of the two types of subjects to construct the dynamic brain functional network（DBFN）.Then we conduct multiple researches on DBFN combined with the feature learning and classification methods in machine learning.The main content of this paper as follows.（1）Extracting sub-networks of early MCI（e MCI）using graph regularized nonnegative matrix factorization（GNMF）.This method aims to reflect the operating principle of integration and separation of brain functions,and this method also overcomes the related shortcomings of using matrix factorization on the aggregate matrix.First,we use the sliding window method to construct DBFNs,and we vectorize and aggregate the DBFNs of normal subjects and e MCI subjects into aggregate matrices.Then we use the GNMF algorithm to perform non-negative matrix decomposition on the aggregate matrices of two types of subjects and restore the sub-networks of two types of subjects.The experimental results show that compared with other algorithms based on matrix factorization,this method can more obviously reflect the similarity between the common sub-networks of two types of subjects,as well as the difference between the distinctive sub-networks of two types of subjects.There are also certain differences between the topological properties of two types of subjects’sub-networks.（2）Analysis of brain functional status for e MCI using variational auto-encoder（VAE）.This method aims to reflect the characteristic of the brain activities of normal subjects and e MCI subjects and the transition rules between functional statuses,this method also overcome the shortcomings of using the deep auto-encoder（DAE）to reduce the dimensionality of the aggregate matrix.First,we use the sliding window method to construct DBFN,and we vectorize the DBFN of two types of subjects and aggregate into an aggregation matrix using functional connectivity strength（FCS）.Then we use VAE to reduce the dimensionality of the aggregate matrices of two types of subjects,and use the Gaussian Mixture Model（GMM）clustering algorithm to cluster the hidden variable matrices of two types of subjects after the dimensionality reduction.Finally,we obtain several common functional networks（CFNs）representing different functional statuses.Perform quantitative analysis and visualization on the CFNs with the most and least occurrences of two types of subjects respectively,and then analyze the transition conditions and time attributes of the functional statuses of the two types of subjects.The experimental results show that there are similarities between the most frequent CFNs of two types of subjects,and differences between the least frequent CFNs of two types of subjects.Compared with the methods in previous researches,this method can reflect this characteristic better,and the status change process and differences in time attributes of two types of subjects can better reflect the dynamic characteristics of two types of subjects’brains.（3）Constructing DBFNs via hyper-graph manifold regularization（HMR）for MCI classification.Nowadays,the BFN construction method based on regularization has a mature framework and has been widely used in the classification and diagnosis of brain diseases.Among them,the BFN construction method based on manifold regularization（MR）has also begun to be applied,but this method only describe the pairwise relationship between two brain regions,and cannot reflect the interaction between multiple brain regions,that is,high-order relationship.Therefore,we first transform the DBFN construction method based on Pearson’s correlation（PC）into an optimization problem,and then we construct the hyper-graph and the hyper-graph manifold regularizer according to DBFN.We introduce the hyper-graph manifold regularizer and the L₁-norm regularizer into the PC-based optimization problem,finally we obtain the final DBFN.We classify normal subjects and MCI subjects to verify the effectiveness of this method.The experimental results show that this method introduces the structure information of the hyper-graph and other prior information into the PC-based optimization problem,so as to restrict the construction of DBFN and effectively improve the classification performance.Researches in this paper have explained the pathogenic mechanism of MCI and the differences between BFNs of two types of subjects from different aspects,and methods involved in this research also provide new ideas for the auxiliary diagnosis of MCI.