Research Of Functional Brain Network Construction And Feature Extraction Method

The functional brain network(FBN)based on resting-state f MRI has been widely used in the early diagnosis of neuropsychiatric diseases and the mining of related physiological markers,which has been a hot research topic in the field of medical image processing.At present,a large number of FBN construction methods and feature extraction methods have been proposed one after another.However,due to the high complexity of the brain and the difficulty of sample collection,there are still many challenges in how to construct a reasonable FBN and extract features with strong discriminative power.In this thesis,the following work has been carried out in the two aspects of FBN construction and feature extraction:(1)In the aspect of FBN construction,a multivariate normal distribution(MVND)brain network construction method based on grouping is proposed.First,all regions of interest are divided into different groups,and then the brain network time series are generated by sliding window technology in each group.It is assumed that the brain network time series obey a certain MVND model and parameter estimation is performed to obtain intra-group low-and high-order brain networks.Further,based on the rs-f MRI signal time series of each group,brain network time series generation and MVND model parameter estimation are carried out to obtain the inter-group low-and high-order brain networks to make up for the feature loss due to grouping.For high-order FBN,this thesis adopts the Kronecker product decomposition to reduce its dimension.This method has the following advantages: 1)It can synchronously construct low-and high-order brain networks with a clear mathematical interpretation.2)It can effectively alleviate the "high-dimensional small sample" problem existing in the whole-brain-based MVND brain network construction method,thereby making the extracted features more discriminative.In this thesis,the method is explored by using the division of medical sub-networks and the clustering of rs-f MRI signal time series as the grouping methods.(2)In terms of feature extraction,a dynamic brain network(DBN)feature extraction method based on two-dimensional principal component analysis(2DPCA)is proposed.Firstly,the DBN is generated by the sliding window technique,and then the DBN is transformed in a certain way to obtain a two-dimensional DBN image.Finally,the principal components of the image are extracted by 2DPCA method,and the obtained principal components are the features.The advantages of this method are: 1)It can effectively alleviate the problems of "high-dimensional small samples" and "timing mismatch" in DBN.2)It can capture valuable structural information of DBN,which reflects the functional linkage information of paired regions of interest.As a three-dimensional DBN transformed into a two-dimensional DBN image from different angles,the functional linkage information of the paired regions of interest that can be retained is different.Therefore,this thesis conducts research on DBN feature extraction based on 2DPCA from the perspective of temporal information and spatial information.The two methods proposed in this thesis both show excellent performance in the classification experiments of patients with autism spectrum disorder and healthy controls,and their results are better than the related control methods,so their effectiveness are verified.