Removing spikes

Excessive head movement, global signal spikes, and uncorrectable artifacts are three popular reasons why you might want to throw out a whole volume from your data, while keeping the surrounding data. We refer to these unhelpful volumes generically as spikes.

Because fMRI data are temporally autocorrelated, you can't just snip the bad volumes out. You also can't just interpolate them, although that wouldn't be too far off the mark. The best way to do this is generally to model the bad volumes out - for each bad volume, to include a covariate that has one value at the time point of interest and another value everywhere else (usually with a mean of zero). This ensures that the bad volumes will not be counted towards your degrees of freedom, and that they will not influence the weights of the non-spike covariates.

It's also important in preprocessing to make sure that the bad volumes don't contaminate the rest of your data. Most preprocessing is spatial, so there's no worry. But slice timing correction involves resampling of your data temporally. One bad volume can end up smeared into adjacent volumes, and depending on how the resampling is done, can end up looking like a complex signal pattern.

To avoid this kind of contamination, it's best to interpolate the bad volumes before the rest of your preprocessing. In VoxBo, you can use the vbinterpolate executable to do this for you, assuming you know which volumes are bad.