ICA Denoising for Event-Related fMRI Studies

The poor SNR of fMRI data requires that many repetitive trials be performed during an event-related experiment to obtain statistically significant levels of inferred brain activity. This is costly in terms of scanner time, necessitates that subjects perform the behavioural task(s) for long durations which may induce fatigue, and vastly increases the amount of data generated. In this paper, we present a method to enhance the statistical effect size using ICA, so that the same level of significance can be obtained with shorter scanning times. We perform ICA on fMRI data from a simple event-related motor task by projecting the original data onto the linear subspace defined by the task-related ICA components. This essentially denoises the signal and results in significant improvement in the effect size. Using simulations we demonstrate that the proposed ICA-denoising procedure is robust to a variety of realistic noise models and enhances the performance of least squares estimates of the evoked hemodynamic response