Hi Regi,

Good question! I think the idea is that we want to match the smoothing that is present in the functional image. In other words, the segmentation is done on the structural image. This will give you a CSF image, next we want to extract signal for this mask from the functional data. To achieve this we first register the CSF image to the preprocessed functional image. However, because we applied smoothing during preprocssing of the functional image, its CSF (or white matter) boundaries will be smoothed. To obtain the same kind of smoothness in our CSF image we apply smoothing to these images.

Next we register the smoothed CSF image to MNI space and calculate the overlap with tissue priors. The result are CSF and white matter masks where we can say with 'great' certainty that they are CSF and white matter in that subject/across subjects. Finally, we regress these masks back to subject space and extract their associated timeseries for use in the nuisance regression.

I hope this clears things up a bit. Of course I would love to hear if you agree/disagree.

Best,
Maarten