Reducing correlated noise in fMRI data.

The sensitivity of functional MRI (fMRI) in detecting neuronal activation is dependent on the relative levels of signal and noise in the time-series data. The temporal noise level within a single voxel is generally substantially higher than the intrinsic NMR (thermal) noise, and the noise is often correlated between voxels. This work introduces and evaluates a method that allows fMRI sensitivity improvement by reduction of these correlated noise sources. The method allows model-free estimation of the correlated noise from brain regions not activated by the functional paradigm using a short (1-2 min) reference scan. A single regressor representing this noise-source estimate is added to the design matrix used in the data analysis. Results obtained from five volunteers show an average t-score improvement of 11.3% and a 24.2% increase in the size of the activated area.