Denoising Task-Based fMRI data: a comparison between methods.

Abstract

Over the last decade BOLD-response related fMRI has seen impressive progress in resolution of the imaging technique. This progress however is limited by the degree to which fMRI can detect localised hemodynamic responses linked to activity and disconnect this from noise. Important noise factors that contribute to the BOLD response include respiration and cardiac cycle. In this research a comparison is made between different methods that pertain to the untanglement of signal from noise, i.e. denoising. Different methods are adopted using Principal Component Analysis (GLMDenoise), Independent Component Analysis (FSL Fix & GLM ICA), and a method using motion and physiological parameters (GLM mp-ph). This in an effort to retrieve signal of interest in a High Field (7T) fMRI visual experiment, comparing the effectiveness and implementational ease of each method. Unfortunately, no conclusive evidence could be reported for the methods efficiently and consistently denoising the data. These findings might be caused by methodological rather than neurovascular issues. Another possible cause are non-linearities in the data and suggest Machine Learning implementations might be useful in future research on fMRI and denoising.