Hi All,
Just wanted to add my take on things...
I agree with correcting separate runs separately where possible and
combining post-analysis.
The interpolation does induce extra smoothing but it is cyclic in
the voxel dimensions such that a whole voxel shift creates no extra
interpolation. However, normally the position only varies slightly
from the mean position within a run but can be very different between
runs and the average amount of interpolation *can* be a problem although
this is both data set and *voxel* specific.
As for covariates - they are there to correct for artefacts introduced
in the scanning my the fact that the subject moved. Hence, even if
perfect motion estimation and correction (in orientation/position)
was done there would *still* be motion induced intensity changes due
to things like spin history, b0 warping, etc...
Putting motion parameters in as regressors is just a very simplistic
way of assuming that these artefacts will bear some linear relation
to the motion parameters. It simply defines a subspace and removes
all signal within this subspace. Problem with that is that can also
remove actual activation signals. Also, this defines a 6 dimensional
subspace whereas the motion of a particular voxel is a specific
combination of these and is therefore 1 dimensional, but then each voxel
needs a separate regressor which, in practice, is not catered for.
With a 6D subspace you get rid of the stuff you want, but also other
things - and this is one reason why it is better to actually apply the
motion correction rather than just try to regress out the major change.
However, there are better ways of dealing with motion artefacts and we are
working on this. Modelling the physics is the key so watch this space...
All the best
Mark
|