Both Marks, thank you for your insights. You answered to the point all my
"motion" questions.
Regards.
Estephan
On Fri, 13 Nov 2009 13:16:11 +0000, Mark Jenkinson <[log in to unmask]> wrote:
>Hi,
>
>Just to add one thing to this - there is also a tool called
>fsl_motion_outliers which will look for timepoints in the
>image where there seems to be excessive motion or
>artifact (based on outliers in the sum of squared changes
>after the motion correction). It will then create a confound
>matrix which will remove the effect of these volumes/timepoints
>from the analysis in the way that Mark described (one EV for
>each timepoint that needs to be removed, with zeros in the
>EV except for a one at just that point in time). You can then
>import this confound matrix using the "Add additional confound EVs"
>in the Stats tab of FEAT.
>
>This would be the recommended way of dealing with quick, sharp
>motions, even when they are quite large. However, make sure
>that if you do have these that the remaining volumes are well
>aligned after the motion correction, as sometimes large motions
>like this can have a knock-on effect in the motion correction, as it
>is done serially.
>
>Hope this helps.
>All the best,
> Mark
>
>On 13 Nov 2009, at 11:49, Mark Woolrich wrote:
>
>> Hi Estephan,
>>
>>> Hello, I have some block-design functional runs (3mm isotropic
>>> voxels) with
>>> varied peaks of motion. Most of them have a single-TR peak of
>>> 4-6mm, while a
>>> few had peaks of 12-20mm. As I understand one have two options to
>>> try to
>>> correct head motion in FSL: a) feeding the motion parameters from the
>>> prestats in the GLM model of FEAT; and b) using melodic to find the
>>> extraneous components that are not task-related.
>>>
>>> I am using as a rule-of-thumb that motions beyond the voxel
>>> dimension (3mm)
>>> or that are time-correlated with the stimulus are to be rejected from
>>> further analysis. As I am trying to salvage some of those runs that
>>> do not
>>> follow this rule, some questions came to mind:
>>>
>>> 1) How can one tell if feeding the motion parameters into GLM
>>> succeeded in
>>> removing the motion-related blobs? Are looking the blobs in the
>>> FEAT output
>>> or in fslview the only options - which are fairly subjective?
>>
>> Putting in motion parameters into the GLM is a good idea but there
>> is no easy way to predict or interpret the effect it has. There can
>> be two intermingled effects it has on the results, either:
>> 1) z-stats go down: due to your motion regressors correctly stopping
>> your effects of interest from erroneously modelling motion that is
>> stimulus correlated.
>> 2) z-stats go up: due to explaining noise variance in the data due
>> to motion that is not stimulus correlated.
>>
>>>
>>> 2) Should I use both methods in conjunction (motion parameters in
>>> GLM+melodic) to obtain a more robust motion correction?
>>
>> No reason why not.
>>
>>>
>>> 3) What are the limits of using these methods; would they be able to
>>> compensate a 12 or even a 20mm motion peak?
>>
>> Hard to give a general rule of thumb. Motion greater than 6mm is
>> getting pretty extreme to handle. Also depends on the type of
>> motion. Short sharp motion within a TR is more problematic than slow
>> drifting motion.
>>
>>>
>>> 4) Motions that are time-locked with stimulus presentation can be
>>> safely
>>> removed using the above methods?
>>
>> Yes - using the motion parameters as confounds will (as mentioned
>> above) stop effects of interest from erroneously modelling motion
>> that is stimulus correlated. Of course one is implicitly relying on
>> the motion correction having worked well, so if you a group with
>> large motion you;ll want to keep a qualitative eye on that as well.
>>
>>>
>>> 5) If only one epoch (stimulus+rest) is affected by the peak motion,
>>> removing the TRs composing the affected epoch using flsmaths would
>>> be an
>>> acceptable approach?
>>>
>>
>> Yes. Although it is easier to leave the data intact and to use a
>> dummy custom EV with all 0s except for 1s in those TRs you want to
>> "remove". Make sure that for the dummy EV you switch off
>> convolution , temporal filtering and temporal derivative.
>>
>>> If anybody has other methods to correct head motion or ways to tell
>>> if the
>>> correction was successful, I would love to hear about it.
>>>
>>> Thanks.
>>>
>>> Estephan Moana, DDS
>>> Graduate Student
>>> Oral Biology PhD Program - Neurobiology track
>>> School of Dentistry, UNC- Chapel Hill
>>>
>>
|