Random noise suppression in fMRI time-series using modified spectral subtraction

We present a novel method for random noise-suppression in functional magnetic resonance imaging (fMRI) time-series based on modified spectral subtraction. The method estimates the signal and noise models at every voxel in the functional data from a small neighborhood, without prior knowledge of the signal characteristics. Spectral subtraction is then performed to obtain a noise-suppressed power spectrum of the voxel under consideration. We demonstrate the performance of the proposed method by preprocessing synthetic as well as real fMRI data. The method is found to efficiently reduce random noise while preserving the deterministic components of the signal, thus, enhancing the sensitivity of the fMRI analysis.