Dear Torsten,
I'm afraid I have bad news.
There is no good and reliable way of determining when to do this automatically.
I have seen cases with small motions that were really problematic, due to
the nature of the motion, the SNR, the correlation with stimulus, the quality of
the scans, etc, etc. I have also seen the opposite - large motions that actually
didn't cause any big problems.
So I never give a rule of thumb like this.
You just need to look carefully at your data and your results to see if the motion
is likely to be causing problems, typically introducing motion-related activation
around the large intensity edges or other non-biologically-plausible activation
patterns.
As for the timeseries fits - are you sure they are from the same voxel? If you
just look at FEAT outputs it will select the best fitting voxel in each case and
you can't compare across different voxels sensibly. If you have a correlation
between the motion and the stimulus then it is also very likely that the statistics
could be "better" (i.e. higher, but in reality biased) without putting in the confound
matrix. So that isn't a reliable test either. If you see that within the same voxel
you get better fits, in terms of lower residuals, without the confound matrix then
this is very strange and would indicate an error somewhere, but I suspect this
is not the case, so have another careful look.
All the best,
Mark
On 11 Aug 2011, at 09:30, Torsten Ruest wrote:
> Hi Mark,
>
> in which cases would you apply the fsl_motion_outliers script? When I inspect the displacement parameters on the report page, the displacements are not really that large. I learned from several sources that displacements < 3 mm and mean absolute displacement less than 0.3 - 0.5 mm are generally considered acceptable, above which scans may probably better be excluded. However if I look at our data, I see spin history effects associated with only little head movement (< 1 mm). Ideally what I 'd like to establish is some kind of rule, like if there is more than 3 mm displacement in one scan, I would run the script. In addition, if there are spin history effects irrespective of large displacements, I'd apply it as well. Would that make sense? The idea is to save as much data as possible, even if that means to be a bit too conservative.
>
> Next I wanted to ask how I would check whether the repair did what I intended to do. I ran one participant with little movement with and without adding the confound matrix. If I am reading the timeseries graph in fslview properly, I see a better model fit in the data without the confound matrix (blue line) compared to including the confound matrix in an identified volume, where the latter looks more like if there's a worse fit in that particular volume, i.e. the fitted (blue) line follows the red line (unfitted) for that volume. Note that there are no visible spin history artifacts in that volume, so I may have actually unnecessarily removed a volume causing a worse fit. So in this particular case I would go without the confound matrix.
>
> I hope that all that I am describing here makes sense.
>
> Looking forward to hear your opinion.
>
> Thanks very much in advance,
>
> Torsten
>
|