An efficient approach to characterize spatio-temporal dependence in cortical surface fMRI data

Abstract

Functional magnetic resonance imaging (fMRI) is a neuroimaging technique known for its ability to capture brain activity non-invasively and at fine spatial resolution (2-3mm). Cortical surface fMRI (cs-fMRI) is a recent development of fMRI that focuses on signals from tissues that have neuronal activities, as opposed to the whole brain. cs-fMRI data is plagued with non-stationary spatial correlations and long temporal dependence which, if inadequately accounted for, can hinder downstream statistical analyses. We propose a fully integrated approach that captures both spatial non-stationarity and varying ranges of temporal dependence across regions of interest. More specifically, we impose non-stationary spatial priors on the latent activation fields and model temporal dependence via fractional Gaussian errors of varying Hurst parameters, which can be studied through a wavelet transformation and its coefficients' variances at different scales. We demonstrate the performance of our proposed approach through simulations and an application to a visual working memory task cs-fMRI dataset.