Research On Construction And Analysis Of Resting-state Brain Network Based On Dual Temporal And Spatial Sparse Representation

The human brain contains more than 10 billion interconnected neurons.It is the high-level central neural system for environment perception,body movement control and consciousness generation.In the past few decades,the structure and function of the brain has been successfully studied by interdisciplinary researchers from cognitive neuroscience,anatomy and information science.However,a full understanding of the functional mechanism of the brain is still a challenging scientific problem due to the numerous neurons and complex structure of the brain.The rapid development of medical imaging technologies in recent years provides possible solutions to the abovementioned problem.Functional magnetic resonance imaging(f MRI)is a non-invasive medical imaging technology with superior temporal and spatial resolution.Constructing the brain network from f MRI data can help to understand the functional mechanism of brain.However,f MRI data are characterized by high dimensionality,high noise and small sample,which brings some problems in construction and analysis of the brain network,including unsatisfactory robustness of the network constructing method,difficulties in qualitative analysis and identification of the subject-specific brain networks,limitation of the previous similarity indexes in quantifying the reproducibility of constructed brain networks.Therefore,the in-depth study on the method for constructing and analyzing resting-state brain network is of theoretical and practical significance.The present work,supported by the National Natural Science Foundation of China(81571341)and the Hunan Provincial Innovation Foundation for Postgraduate(CX2016B127),aims to construct and analyze brain networks using resting-state f MRI data.First,a novel method with dual temporal and spatial sparse representation(DTSSR)is proposed to construct the subject-specific brain networks and group-level intrinsic connectivity networks(ICNs).Then,a multivariate correlation coefficient is proposed to quantify the retest reproducibility of the group-level ICNs obtained by DTSSR.Finally,the proposed DTSSR is applied to analyze resting-state f MRI data of adult patients with attention deficit hyperactivity disorder(ADHD).The main contributions and conclusions of the present work are as follows:1.The sparse representation method for constructing subject-specific brain networks from resting state f MRI is studied.Traditional Independent Component Analysis(ICA)method assumes that the source components of f MRI data are statistically independent,which lacks neurological and physiological basis.The sparse representation algorithm originating from neurophysiological research,is therefore applied in the present work to construct subject-specific brain network from resting-state f MRI data.Firstly,the resting-state f MRI raw data are preprocessed to improve the signal-to-noise ratio.Then,sparse representation algorithm with online dictionary learning is used to decompose the preprocessed individual f MRI signal matrix,thereby obtaining the subject-specific brain networks.Finally,the brain networks reported in literatures are applied to match and analyze the obtained subject-specific brain networks.Experimental results on the public resting-state f MRI dataset Leiden?2180 show that the correlation coefficients between the matched subject-specific brain networks and the reported brain networks are 0.24-0.76,and the P-values on the test of significance are less than 10^-5,indicating that sparse representation can effectively construct subject-specific brain networks from the resting-state f MRI dataset.2.A novel dual temporal and spatial sparse representation(DTSSR)is proposed to construct group-level brain networks.To solve the difficulty in identification and qualitative analysis of the subject-specific brain networks,a simple mathematical model is proposed to describe the complex nonlinear dynamic coupling mechanism of intrinsic connectivity networks(ICN),according to the characteristics of hierarchical coupling mechanism of brain networks.Based on the model,DTSSR is proposed to construct brain network from resting-state f MRI data,with which the group-level ICNs with group commonalities as well as the coupling parameter matrix are obtained.The proposed DTSSR method was tested on Leiden?2180,Leiden?2200 and our own dataset.Experimental results show that:the proposed DTSSR is advantage in identification and qualitative analysis of the subject-specific brain networks;the obtained group-level ICNs are interpretable by current brain science knowledge;DTSSR has better robustness when compared with ICA-based method.3.A novel evaluation index is proposed to quantify the retest reproducibility of brain network.Traditional similarity indexes are limited in directly quantifying retest reproducibility of data by three or more measurement.To solve this problem,a novel multivariate correlation coefficient,which is based on zero-mean normalization and intraclass correlation coefficient(Z-ICC),is proposed to quantify retest reproducibility of the group-level ICN sets obtained from f MRI data scanned many times.Experiments on the public retest f MRI dataset NYU?TRT demonstrate that:(i)the proposed Z-ICC index is effective in quantifying the retest reproducibility of the group-level ICNs that are obtained from f MRI datasets scanned three times;(ii)compared with other indexes,the obtained results using Z-ICC is more consistent with visual inspection;(iii)compared with the traditional TC-GICA method,the proposed DTSSR method achieves better retest reproducibility.4.The proposed DTSSR is applied in the analysis of the resting-state f MRI data of adult patients with attention deficit hyperactivity disorder(ADHD).The ADHD?ICNs are obtained,based on which a computer-aided diagnosis method for ADHD is designed,using Support Vector Machine as the classifier.Experiments on the public dataset New York?a?ADHD show that attention network and execution control network are related with ADHD,which involve attention control,behavioral inhibition and emotion control.These obtained two networks can be verified by previous studies in brain science.Experiments on public datasets New York?a?ADHD(ADHD group)and New York?a(health control group)show that the proposed method achieves a classification accuracy rate of 97.9%,which provides valuable reference for clinical diagnosis of ADHD.