Analysis Of Dynamic Brain Function Connection Network Based On Tensor Decomposition

The rapid development of brain imaging technology has promoted human exploration of the field of brain science and cognition.Similarly,methods such as machine learning and data mining have important theoretical basis and application value for the analysis of brain image data.The Functional Connectivity(FC)network estimated from the time series of brain images provides a powerful tool for analyzing the functional structure of the human brain in health,disease,and developmental states.The FC model estimates of most studies in the past are based on the assumption of the stability of the entire scan time and space.However,the actual FC shows time volatility,which means that it is too simple to analyze the stability scan in a complete time,and it is impossible to capture the entire process of the entire activity.Combined with the analysis of the dynamic performance requirements of the FC network,using resting brain image data,a tensor model is used to analyze the dynamic FC network of people of different ages.The main research contents are:(1)Aiming at the problem of FC volatility with time,the dynamic FC network is analyzed based on a tensor Tucker decomposition algorithm.This method is mainly based on the two-step method to track the dynamic connection of the whole brain.Specifically,the result of tensor decomposition is used to divide the state of the time-varying network using the dynamic connection detection method.Finally,the topology of the FC network is captured based on the tensor method,thereby providing accurate change point detection and status summary.(2)Due to the high computational complexity of Tucker decomposition and the inability to quickly integrate analysis in multiple dimensions,a dynamic FC analysis based on sparse tensor CP decomposition algorithm is designed.For the young group and the children group,first use the resting state data to construct a tensor model based on the different information of each dimension of the tensor,and optimize the algorithm.Then use sliding window method and dynamic connection detection method to construct dynamic network.Finally,the two network construction methods are evaluated on different simulated data and real data.The results show that the two methods of network construction can capture the dynamics of FC,and the dynamic connection detection method overcomes the limitation of the window size ofthe sliding window method.(3)Aiming at the problem of small samples with multiple features,the corresponding punishment constraints are improved on the tensor model.First,based on the SVM and SVR methods,the decomposition results are used to realize the functions of age classification and prediction,and the performance of the original tensor algorithm and other decomposition algorithms and improved algorithms are evaluated.The results show that the improved algorithm is superior to other algorithms and improves the model performance.Then compare and analyze the improved tensor decomposition algorithm and the static and dynamic FC constructed by the two network construction methods.Finally,the connection strength of brain function areas in different states of the young group and the child group was analyzed.The experimental results show that the dynamic connection detection algorithm based on tensor CP decomposition is the most accurate for dynamic FC analysis and can obtain more connection information of different brain regions.