Link to paper
The full paper is available here.
You can also find the paper on PapersWithCode here.
Abstract
- Many fMRI studies rely on estimates of brain networks.
- Existing methods for estimating the community structure of networks lack validation.
- A new multilayer community detection method based on Bayesian latent block modelling is developed.
- The method can detect the group-level community structure of weighted functional networks.
- A new community structure-based multivariate Gaussian generative model is proposed for validation.
- The method is tested using real working memory task-fMRI data from the Human Connectome Project.
Paper Content
Introduction
- Brain function or cognition can be described in terms of multiscale hierarchical organization.
- Functional connectivity is a measure of statistical dependence between regional time series.
- Changes in cognitive states and noise affect the transient functional interaction between pairs of nodes.
- Single-subject estimates of functional interactions between brain regions can be unreliable.
- Community-detection methods have been used to characterize brain states in functional networks.
- Recent studies have focused on multiple networks from different subjects to reduce uncertainty.
- Mean-field models can be used as a forward model which can be inverted given the empirical data.
- A new method based on hierarchical Bayesian modelling was proposed to capture the variability between subjects.
- The method uses latent block modelling and MCMC sampling with absorption and ejection strategy.
- The method was validated by simulating fMRI data with a multivariate Gaussian generative model.
- The method was applied to working memory task-fMRI data from the Human Connectome Project.
Methods
- Developed a novel multilayer community detection method
- Method based on hierarchical Bayesian modelling
- Estimates community memberships and number of communities of functional brain networks
- Uses MCMC strategy to estimate individual-level community memberships
- Uses conjugate Bayesian pairs to model estimated individual-level community memberships and individual adjacency matrices
- Performs parameter inference to depict group representative network
Individual-level modelling of the community structure
- Latent block model (LBM) used to model weighted adjacency matrix measuring functional connectivity (FC) of a brain state
- Nodes assigned to K communities, K is a random variable following a Poisson distribution
- Community memberships denoted as latent label vector z, each z i independently follows a categorical distribution
- Model parameter ฯ kl follows a Normal-Inverse-Gamma (NIG) prior
- Estimate community structure at individual level by sampling latent label vector z from collapsed posterior distribution
- Update latent label vector z using Gibbs move, M3 move, absorption move, and ejection move
- Group-level modelling to analyze group representative network from all subjects
- Community memberships treated as realizations of a latent group-level community structure
- Bayesian model with categorical-Dirichlet pair used to fit community memberships
- Maximum label assignment probability matrix (MLAPM) calculated by maximizing a label assignment probability matrix (LAPM)
- Model functional connectivity (FC) of discrete brain states using Normal-NIG conjugate pair
- Mean and variance connectivity at group level estimated by drawing samples from Normal (N) and Inverse Gamma (IG) posterior densities
Group-level modelling of the community structure and connectivity
- Modeled community memberships estimated from individual-level analysis
- Defined matrix to represent latent labels for all subjects
- Modeled for S subjects for node i using categorical-Dirichlet conjugate pair
- Used maximum LAPM (MLAPM) to provide information of community structure
- Defined connectivity parameter for element ij and Gaussian distribution for connectivity of each subject
- Calculated posterior distribution for group connectivity as NIG distribution
Experiments
Generative modelling and synthetic fmri data experiments
- Used hierarchical Bayesian modelling method to simulate synthetic fMRI data
- Generated D segments of Gaussian time series from D different block covariance matrices
- Nodes assigned to K true communities, whose assignments differ in different segments
- Generated true label vectors to determine form of covariance matrices
- Simulated time series data with desired underlying community structure
- Convolved multivariate Gaussian data with canonical HRF
- Generated 100 instances of synthetic fMRI data with N = 35 nodes and T = 180 time points
- Generated D = 7 data segments with length of 20 frames for each segment
- Set numbers of communities in data segments to K true = {3, 4, 5, 3, 5, 4, 3}
Task fmri data experiments
- 100 unrelated healthy subjects participated in the study
- Data was collected using a 3T Siemens Skyra
- Participants completed a N-back working memory task
- Data was preprocessed using FSL
- GLM analysis was used to select brain regions of interest
- Cluster-wise inference was applied with a CDT of Z = 3.1
- A family-wise error-corrected threshold of P = 0.05 was set
- A binary spherical mask with a 6 mm radius was created
- Time series were extracted from 35 brain regions of interest
Results
- Validated multilayer community detection using synthetic and real fMRI data
- Used Bayesian change-point detection to identify brain states
- Used hierarchical Bayesian modelling to fit FC and estimate underlying community structure
Method validation using synthetic fmri data
- Used synthetic fMRI data with SNR of 10dB and 5dB
- Estimated functional networks for five states between change-points
- Used MCMC allocation sampler to estimate community memberships
- Used categorical-Dirichlet conjugate pairs to model latent labels
- Used label-switching method to solve label switching problem
- Estimated mean and variance of connectivity between nodes using Normal-NIG conjugate pair
- Found posterior samples of mean and variance were highly correlated with real mean and variance
Method evaluation using working memory task-fmri data
- Extracted time series of 35 nodes from fMRI data
- Analyzed data from left-right phase encoding
- Visualized brain network connectivity using BrainNet Viewer and Circos
- Used relabelling algorithm to reassign labels across different states
- Group-level weighted edges at sparsity level of 10%
- Nodes clustered into communities with different connectivity densities
- Strong connections in communities k = 2 and 4, weak connections in communities k = 1 and 3
- Estimation of mean and variance connectivity for discrete brain states
- Small changes to latent FC patterns may be sufficient to give rise to a wide variety of cognitive and behavioural states
- Group-level modelling of functional connectivity and community structure
- Hierarchical Bayesian modelling to characterize mean and variance of weighted FC
- Relaxed assumption of fixing number of communities K
- Absorption and ejection moves integrated into MCMC sampler
- Validated group-level community memberships and connectivity estimates using synthetic fMRI data
- Estimation of community memberships at individual level using working memory task-fMRI data
- Estimation of group-level FC using hierarchical Bayesian modelling
- Future research to relate HCP behavioural data to FC
- Future research to relate community detection and node metadata to network structure