Hi - what Mark says makes a lot of sense.

I would just add a couple of minor points:

- You should definitely make sure you're using FSL 4.1 as it has more  
liberal masking than 4.0 (for these very reasons).

- Not sure what Vince is thinking as the spatial smoothing is  
effectively done at the same stage in FSL and SPM, AFAIK.

- But yes, reducing the masking threshold will definitely make the  
masking more liberal. Doing this is fine.

Cheers.





On 14 May 2009, at 08:49, Mark Jenkinson wrote:

> Hi Stefan,
>
> The problem has no easy answer I'm afraid.
> It isn't something that we've really come across very much, although
> we have been thinking about ways of better incorporating this in
> future releases now.
>
> There are three potential things that you could do.
>
> One is to do what David suggests and just make the masks bigger (e.g.
> by lowering the %brain/background threshold) but this may include
> voxels that do not contain useful signal and hence bias your results,
> although with such a high N, this might not be too bad.
>
> The second thing to do is to include some voxel-specific regressors
> which model out the effects of the missing data points.  You would
> also need to increase the masks to get these voxels included, and
> you would need one regressor per number of subjects that you wanted
> to reinstate.  For example, to include all voxels where one subject
> only was missing you would need one voxel-wise regressor which
> contained zeros for all timepoints except a single one where there
> was missing data.  This doesn't have to be the same subject in each
> voxel - this is the advantage of the voxel-wise regressors - so that
> you would have certain voxels containing an all zero regressor (if
> they had data for all timepoints) and others with a single one in the
> appropriate slot corresponding to the subject which was missing
> *for that voxel*.  Note that timepoints = subject index in the above.
> This is a nicer mathematical solution but is practically a little  
> difficult
> to do and if you wanted to include voxels where two subjects were
> missing then you'd need another voxel-wise regressor and so on.
> Therefore if you wanted to include voxels where you had N=300
> or 299 or 298 or ... 270, then you'd need to make 30 voxel-wise
> regressors and include them all which is cumbersome and may
> slow the analysis down a fair bit.
>
> The third solution would be to use the outlier detection mechanism
> to do the work for you.  This again requires enlarging the masks for
> each subject (to include all the voxels you want to - and it may be
> better to do this by replacing the mask files in reg_standard directly
> with a common mask that you want).  Once you've done this you
> would need to replace values in the cope images (again in  
> reg_standard)
> which were previously outside the mask (and hence would currently
> be zero) with values which are clearly outliers.  You should be able
> to work out a sensible outlier value by looking at your current
> analyses and seeing how much the valid copes vary.  If you do this
> the outlier detection should hopefully identify these subjects as
> outliers in those voxels and hence effectively remove them from
> the analysis.  I must say that I haven't tried this solution (or in  
> fact
> the previous one) but in principle it should work.
>
> Sorry I haven't got a nicer and easier solution for you.
>
> All the best,
> 	Mark
>
>
> On 14 May 2009, at 04:10, David V. Smith wrote:
>
>> Hi Stefan,
>>
>> Which version of FSL are you using? I believe 4.0 was a bit too  
>> conservative when generating the masks (see previous posts on this  
>> matter), but the newer versions don’t really have a problem with  
>> this.
>>
>> But, regardless of what version you’re using, once a voxel is gone,  
>> it’s gone for everyone. So, even if you have just one bad mask in  
>> your dataset, it will mess up the others.
>>
>> I think your approach to overcoming this (i.e., lowering the %brain/ 
>> background threshold) is fine. Just make sure you do it for everyone.
>>
>> Hope this helps,
>> David
>>
>>
>> --
>> David V. Smith
>> Graduate Student, Huettel Lab
>> Center for Cognitive Neuroscience
>> Duke University
>> Box 90999
>> Durham, NC 27708
>> Lab phone: (919) 668-3635
>> Lab website: http://www.duke.edu/web/mind/level2/faculty/huettel/
>>
>> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On  
>> Behalf Of Stefan Ehrlich
>> Sent: Wednesday, May 13, 2009 7:52 PM
>> To: [log in to unmask]
>> Subject: [FSL] problems due to concatenated masks in GFEAT of a  
>> very large study - 3rd posting - no answer?
>>
>> Dear FSL'ers,
>>
>>
>>
>> I had posted that earlier. This is the 3rd posting. It would be  
>> *very* helpful if someone could give me some advice on that. Please  
>> also let me also know If the problem is not described clearly  
>> enough or if there is no easy answer.
>>
>>
>>
>> I am running FEAT on a large fMRI dataset (n > 300). Each subject  
>> has 3 runs consisting of 16 blocks of a memory paradigm. After  
>> hundreds of hours of computing I have processed all first- and  
>> second level-analyses (from here on referred to as cross-runs). I  
>> have checked the registrations and individual as well as cross-run  
>> activation-maps (retrieval versus fixation) for a subset of  
>> subjects and all seems fine. As a next step I ran a “Single-Group  
>> Average” over all 300 cross-runs just to get an impression of the  
>> overall activation patterns. The results were in line with previous  
>> studies but in the top slices of the brain (horizontal slices) as  
>> well as in a rim covering the parts of the brain closest to the  
>> skull there was no activation whatsoever. I found out that this was  
>> probably due to the mask.nii of the gfeat. This group mask did not  
>> include the top slices and the outer rim .
>>
>> Subsequently I went back and checked all cross-run mask.nii and  
>> identified a very few masks which were missing several top slices  
>> (due to bad positioning in the scanner, I guess). After deleting  
>> these subjects from the overall analysis my final results and the  
>> “Single-Group Average” mask looks much better. However, the outer  
>> rim is still missing.
>>
>> It seems like FEAT concatenates all cross-run masks and does not  
>> include voxels which have a missing value in any single mask. Vince  
>> Calhoun told me on the phone that this might be due to the fact  
>> that FSL smoothes relatively late in the processing stream (in  
>> contrast to SPM). Concequently I went back and changed the brain/ 
>> background threshold (brain_thresh) from 10 to 1 and rerun a few  
>> subjects which had slightly impaired cross-run masks. With the new  
>> threshold more voxels get included. Do you think that is an  
>> approbriate approach? Has anybody experienced that problem before?  
>> Are there other solutions?
>>
>>
>>
>> Thank you so much for your thoughts!
>>
>>
>>
>> Stefan
>>
>>
>>
>>
>


---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director,  Oxford University FMRIB Centre

FMRIB, JR Hospital, Headington, Oxford  OX3 9DU, UK
+44 (0) 1865 222726  (fax 222717)
[log in to unmask]    http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------