Head movement between conditions can be a very bad thing, but the best way
forward would depend on many things. The main problem is that if you try to
model the movement in the design matrix, then much of any real activation
signals may be modelled out. If you don't model movement, then you will
detect significant differences among your data, but you are not in a position
to say whether the differences are due to differences in brain activity or
movement. Even if you do model out motion, it would be unlikely that all
possible motion artifacts would be accounted for - so you still can't be
absolutely certain that significant differences are really due to activity
differences.
1) Do you have artifacts in the data (e.g. distortions in the phase encode
direction or ghost artifacts)? These can not all be corrected by rigid-body
transforming the data, so it may be more important to include estimated
movement parameters as confounds in the design matrix.
2) Are the movements correlated with the experimental design? If there are no
correlations, then all the uncorrected noise-related variance will do is
increase your residual variance, but shouldn't give you any artificial
activations.
3) Are there sudden movements (sneezes etc)? If so, then it may be an idea to
model these out (by including additional columns in the design matrix
containing all zeros, except for a one in the appropriate row).
4) Are the movements real? Some scanners have been known to produce images
that drift in position over time.
5) How many scans do you have and how strong do you expect the activations to
be? If you have more scans, then you can regress out more variance due to
movement, and still have some signal left. If you only have a few scans,
then you may want to try correcting with the "Realign and Unwarp" option, in
which case you probably wouldn't include the estimated movement parameters in
the design matrix.
Determining the best course of action is an empirical matter, and as I never
actually run the stats, I don't have a good intuition. Perhaps others can
comment.
Best regards,
-John
> I'm also worrying about the head movement between conditions. Our
> experiments have 3 condition A B and C, each runs for 5 minutes. I
> observed for some subjects, the head motions are above 8 mm, so for B-A
> or C-A, it's real troublesome. I tried to realign each run separately
> and then coregister the mean image of each run to the T1 image, and
> followed by normalization, smoothing. But unfortunately, the GLM results
> seemed worse than realigning them together. Does anybody have some good
> ideas?
>
> Thanks.
>
> -----Original Message-----
> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> On Behalf Of Stephen J. Fromm
> Sent: Monday, January 24, 2005 12:25 PM
> To: [log in to unmask]
> Subject: Re: [SPM] movement between runs
>
> On Mon, 24 Jan 2005 16:05:08 +0100, Chiara Begliomini
>
> <[log in to unmask]> wrote:
> >Dear SPM experts,
> >
> >
> >
> >I am pretty new in the fMRI world and I am trying to analize some fMRI
>
> data coming from a grasping experiment (event-related design, with
> variable SOA). Subjects are requested to grasp real objects while being
> scanned, thus they really have to move their right arm and grasp objects
> presented in front of them.
>
> >We usually observe a lot of head motion (more than 6 mm translation,
>
> and
> 2°-3° degrees rotation) but we also noticed that movement occurs mostly
> between runs, during the break between a run and the following one.
> During
> the run, so when subjects are performing the grasping task, motion is
> usually not worse than 1.5 mm for translation and 1° degree rotation.
> How
> could I deal with such a problem? Am I methodologically allowed to
> perform
> the realign for each run separately (like if I were performing the
> realign
> for different subjects) and then do the following steps (coregistration,
> normalisation and smoothing) as usual, considering data coming from the
> realignment step as coming from only one subject (like they really are)?
> If not, which is the best procedure to try to exclude head movement as
> much as possible from data?
>
> FWIW my feeling is that motion between runs isn't as important as motion
> within a run, *IF* subtractions are not primarily "between runs." (Of
> course, really large movement even between runs could be a problem, but
> 6
> mm might be OK.)
>
> For example, if events A and B occur in all runs in a relatively
> balanced
> manner, and you're subtracting B from A, then you can think of this as
> averaging the results of doing a separate subtraction in each run.
>
> If A were mostly in run1 and B were mostly in run2 (which is somewhat
> problematic and unusual for fMRI anyway), then taking A - B "subtracts"
> the runs and inter-run motion becomes a more important consideration.
>
> Assuming what I just said is reasonable, then in terms of realignment,
> if
> the inter-run motion isn't too outlandish (6mm doesn't sound ridiculous,
> though you should visually check quality after the realignment is done),
> then it's probably OK to just realign all the runs together the usual
> way. That's because SPM essentially first aligns the volumes of each
> run
> separately, then aligns the first volumes of runs 2, 3, ... with the
> first
> volume of the first run, and pulls the other volumes along with those.
>
> HTH,
>
> S
>
> >I thank you in advance for your help,
> >
> >
> >
> >Chiara
|