Spatial Bayesian hierarchical modelling of functional Magnetic Resonance Imaging (fMRI) data

Abstract

Functional magnetic resonance imaging (fMRI) is a technique that measures changes in blood oxygenation in the brain as a result of a stimulus. This technique provides insight into the vast hidden structures of the brain. The aim of this study was to obtain activation amplitudes and regions of activation in the brain for the motor tasks: visual cue, left and right hand, left and right foot, and tongue. Cortical surface (cs)-fMRI data was used for the analysis, gathered from 10 adults from the Human Connectome Project (HCP). A spatial Bayesian general linear model (GLM) was implemented to obtain estimates of activation amplitudes at single and multiple subject levels. Two methods were employed for the multiple subject analysis: the joint and the two-level approach. The integrated nested Laplacian approximation (INLA) technique was used for Bayesian computation. The regions of activation were identified with the use of joint posterior probability maps (PPM), thresholded at different levels. The results obtained were illustrated as figures of inflated brains to aid in the visualisation. Verification of the accuracy of the results was achieved by comparing these figures to the literature on the physiology of the brain. The single and multi-subject Bayesians GLMs depicted accurate activation amplitudes. The two-level approach for the group-level analysis produced smoother activation amplitude estimates than the joint approach. Moreover, the threshold level 𝛾 = 1% illustrated the most targeted activations in the regions associated with the tasks. Accuracy in the results and the computational efficiency of the method suggest that using a Bayesian approach to account for spatial dependencies in cs-fMRI motor task studies is advantageous.