Confounds to clean fMRI

Using Confounds to Clean fMRI Data

Means

BOLD Signal in gray matter is interpreted as delayed evidence of neural activity. But there is no interpretation of BOLD signal in white matter or CSF. So signal there is more likely to be artifact, scanner noise, physiological confound, etc. By regressing these nuisance signals out in a GLM, we are likely to remove the noise we don’t care about and leave behind cleaner signal we do.

csf, white_matter, csf_wm

Global signal can be regressed to normalize the data.

global_signal

aCompCors

Anatomical component corrections.

a_comp_cor_*

tCompCors

Temporal component corrections.

tcompcor
t_comp_cor_*

DVARS

Volume-to-volume root mean squared change in BOLD signal averaged over all voxels.

dvars, std_dvars

Motion Signal

The volume-to-volume motion detected during motion correction is saved as vectors in the confounds files. The spatial motion has already been reversed by transforming each volume to match. But the deeper effects of movement on things like spin-history within the BOLD signal remain. Regressing out signal related to this motion can be accomplished by including these vectors as nuisance variables in the GLM.

trans_x, trans_y, trans_z, rot_x, rot_y, rot_z
framewise_displacement, rmsd

Motion Outliers

Motion censoring (or “scrubbing”) (Power, et al. 2011) is using a vector of 0’s, with 1’s representing TRs where high motion occurred, in your GLM to soak up the variance unique to that one TR, removing its effect from your data. fMRIPrep adds these vectors to the end of the confounds.tsv file. Carp, 2011 suggests doing this before bandpass filtering and Power, 2012 describes how that may depend on the nature of the motion.

motion_outlier*

Also

Band-pass filtering, usually from 0.008-0.01Hz on the low end to 0.08-1.0Hz on the high end, is generally accepted to remove non-BOLD signals. Carp, 2011 filtered 0.009-0.08Hz, then blurred with 6mm FWHM kernel.

Satterthwaite excluded subjects with gross motion > 0.55mm mean relative displacement, leaving only those with <=0.20mm. They took six motion confound parameters and condensed them into RMS (root mean square) displacement, then the entire timeseries was condensed into a scalar MRD (mean relative displacement).