Assessing Brain Connectivity Using Functional Magnetic Resonance Imaging

The human brain is a complicated and large network which is composed of about100billion interconnected neurons. The accomplishment of any function of the brainrequires the collaboration and interaction of widely distributed brain regions. Tounderstand the mechanism of the brain more clearly, we need to examine the humanbrain as an integrative network and study brain connectivity. As a new noninvasivemedical imaging technology, functional magnetic resonance imaging (fMRI) has beenused widely in various fields of brain science recently. Focusing on the brainconnectivity analysis methods, this dissertation studies functional and effectiveconnectivity of the brain comprehensively and deeply using fMRI data.Functional connectivity which is the simplest and most direct characterization offunctional interactions is studied. Starting with functional connectivity analysismethods, this dissertation introduces the principles of model-driven methods anddata-driven methods. Resting-state functional connectivity of the default modenetwork in major depressive disorder (MDD) was studied using independentcomponent analysis. In addition, the effect of antidepressants on functionalconnectivity of the default mode network was investigated. The results showed that theresting-state default mode network of the MDD subjects dissociated into twosub-networks and the anterior sub-network exhibited resistance to antidepressants.This finding is of great significance in illuminating the neural mechanisms of majordepression.Functional connectivity measures undirected statistical associations among brainregions and could not make inferences about the directions of brain connections. Inorder to investigate the information flow in the brain network, this dissertation furtherstudies directed brain connectivity－effective connectivity. The basic principles ofdeterministic dynamic causal modeling (DCM) are systematically expounded anddeduced. Furthermore, effective connectivity of the response inhibition network ofpatients with Internet addiction was studied using deterministic DCM, andsignificantly decreased effective connectivity was found in the patients whileperforming a GoStop task. The results may provide new insights into the neuralsubstrates of Internet addiction.A model selection method based on network discovery is proposed. Generally, dueto the current lack of detailed knowledge on anatomical connectivity in the humanbrain, researchers usually define some competing models when performing effectiveconnectivity analysis and use a Bayesian model selection procedure to identify the bestmodel. However, as the number of nodes in the network increases, the model spaceexpanded rapidly. It becomes impossible for the traditional model selection method to find the best model in such huge amount of models. This dissertation proposes a newmodel selection method based on network discovery which can find the best reducedmodel in the whole model space only by estimating parameters of a full connectedmodel. Moreover, the new method does not require researchers to define competingmodels according to their prior knowledge, thus makes DCM become a data-drivenmethod.Stochastic DCM for fMRI is studied. Deterministic DCM treats the brain as adeterministic dynamic system and assumes that changes in neuronal activity are causedonly by exogenous inputs. However, many studies have suggested that neuronalactivity is also driven by endogenous physiological fluctuations in addition toexogenous inputs. If these endogenous fluctuations are left unaccounted, it will lead toincomplete model specification and misleading parameter estimates. This dissertationstudies stochastic DCM for fMRI which allows for endogenous fluctuations in stateequations. Simulated data and fMRI data are used to compare deterministic DCM andstochastic DCM. The results show that parameter estimates of stochastic DCM aremore precise than that of deterministic DCM. Moreover, stochastic DCM breaks thebottleneck of deterministic DCM and can be used to study resting-state effectiveconnectivity. A framework for resting-state effective connectivity analysis is proposedin this dissertation by combing group independent component analysis and stochasticDCM. Effective connectivity of the default mode network at rest and duringperformance of cognitive tasks are studied and compared for the first time. Althoughactivity of the default mode network was decreased, increased effective connectivity ofthis network was found during task performance.Generalised filtering is studied in this dissertation. Inversion of stochastic DCM isa triple estimation problem and established schemes rely on the mean-fieldapproximation which assumes conditional independence among unknown quantities.However, this assumption is usually not tenable in fMRI data analysis. Recently wehave developed a new scheme called generalised filtering which dispenses with themean-field approximation and can provide online parameter estimation. In thisdissertation, we continue to study generalised filtering on the basis of our previouswork. The validity of generalised filtering is examined using simulated data,single-subject and group fMRI data.