Mixed spectrum analysis in spatial context: Application to fMRI

In functional MRI (fMRI), the noise is often considered to have a fixed model of temporal autocorrelation. However, the correlation structure of brain voxels is generally unknown and varies due to the regional effects of neurogenic noises and pulsatile effects. In this paper, we apply mixed spectrum analysis to separate the discrete (task-related) spectrum and continuous (noise-related) spectrum of fMRI signals in task-related experiments, without assuming a fixed model for temporal autocorrelation. The purpose is to improve the detection of activation in periodic task-related fMRI data. We have also modified the M spectral estimator (MSPEC) method, generally applied on individual time-series, in order to incorporate contextual information in fRMI data. The contextual MSPEC (C-MSPEC) considerably reduces computation cost of the MSPEC algorithm. We demonstrate the effectiveness of the C-MSPEC method on both simulated and real task-related fMRI data.