A spatio-temporal nonparametric Bayesian variable selection model of fMRI data for clustering correlated time...

Area of Science:

• Neuroimaging and Computational Neuroscience
• Statistical Modeling and Machine Learning

Background:

• Functional magnetic resonance imaging (fMRI) generates complex spatio-temporal data.
• Analyzing fMRI data requires methods to detect brain activity and understand voxel relationships.
• Existing methods may not jointly address activation detection and voxel time series association.

Purpose of the Study:

• To present a novel wavelet-based Bayesian nonparametric regression model for fMRI data analysis.
• To develop a joint framework for detecting brain regions with neuronal activity and inferring voxel clustering.
• To model voxel-dependent hemodynamic response function (HRF) shapes.

Main Methods:

• Utilized mixture priors with a spike at zero for detecting activated brain regions.
• Employed a Markov random field (MRF) prior to account for spatial correlations among voxels.
• Modeled voxel time series association using correlated errors with long memory properties and discrete wavelet transforms.
• Applied a Dirichlet process (DP) prior for voxel clustering.
• Inferred model parameters using Markov Chain Monte Carlo (MCMC) sampling techniques.

Main Results:

• The proposed model successfully detects brain regions exhibiting neuronal activity in response to stimuli.
• The framework effectively infers associations and clusters spatially remote voxels with similar fMRI time series.
• Performance was validated on both simulated (block- and event-related designs) and real fMRI data.

Conclusions:

• The developed wavelet-based Bayesian nonparametric model offers a unified approach for fMRI analysis.
• This method enhances the ability to simultaneously identify brain activation and uncover functional relationships between brain regions.
• The model's flexibility and performance suggest its utility for advancing neuroscientific research using fMRI.