Font Size: a A A

Dynamic Brain Functional Network Analysis Based On Unsupervised Learning

Posted on:2020-11-11Degree:MasterType:Thesis
Country:ChinaCandidate:Y CuiFull Text:PDF
GTID:2404330578454857Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Resting-state functional magnetic resonance imaging(fMRI)is one of the important techniques for studying the human functional brain.The analysis of the dynamic properties of the functional brain network is a hot topic in the functional brain research,and the division of functional brain network state is an important aspect of functional brain network dynamic attribute analysis.At present,the functional brain network state partitioning strategy widely used in the world is based on k-means clustering algorithm to cluster short-term functional network matrices,while the k-means clustering algorithm has difficulty in determining the number of clusters and the initial clustering center which influences the clustering result significantly.And the k-means clustering algorithm cannot consider the dynamic temporality of the functional brain network.In order to effectively solve these problems,this research has carried out the following three tasks:First,the functional brain network state division based on density peak clustering.Considering the density peak clustering can visually display the reasonable number of categories,which solve the problem effectively that k-values in k-means clustering are difficult to determine,we took the lead in introducing density peak clustering to carry out the study of human brain dynamic functional network partitioning.In addition,considering the high dimension of the functional connection features obtained based on the sliding window method,before the density peak clustering,we used the multidimensional scaling analysis algorithm to effectively reduce the dimension based on multiple distance metrics.The experimental results showed that the number of effective functional brain network state was between 3 and 5.Second,based on the Hidden Markov Model(HMM),the study of functional brain network state transition.Considering that HMM can make full use of the dynamic time series of functional brain network,it can effectively compensate for the shortcomings of k-means clustering ignoring dynamic time series information.In addition,considering the high dimension of the functional connection features obtained based on the sliding window method,before using the HMM,we effectively reduced the dimension of the functional connection features based on various dimensionality reduction algorithms.The experimental results showed that the time-based unsupervised algorithm can obtain the dynamic attribute information of human functional brain network with different perspectives.Third,the functional brain network research based on sequential subspace clustering.Considering the“Spatiotemporal Pattern,repeated in the human brain with time i5 of great significance,we were the first to introduce sequential subspace clustering to carry out dynamic functional brain network partitioning research.In addition,considering the short-term functional network is different from the meaning of the full-order time series data,we used sequential subspace clustering to simultaneously capture the "Spatiotemporal Pattern,existing in the two kinds of data.Both experimental results indicated that there is a strong correlation between functional brain network states and development.The innovation of this research lies in(1)Introducing density peak clustering to solve the problem that the number of categories in the human functional brain network state division based on k-means clustering is difficult to determine;(2)Introducing HMM and fully considering the timing of human functional brain network states;(3)Introducing the concept of sequential subspace clustering,taking full account of the spatiotemporal pattern of the temporal and spatial domains of the human functional brain network.The experimental results showed that the three types of functional brain network state partitioning strategies proposed in this paper can be available and effective.
Keywords/Search Tags:Resting-state fMRI, Density peak cluster, Hidden Markov Model, Sequential subspace clustering, Dynamic functional connectivity
PDF Full Text Request
Related items