Hi Joe,
I had the same problem a few months ago. Using the default FSL
settings, I wound
up losing tons of voxels inside the brain in my thirdlevel analyses. Reducing
the %brain/background setting to one helped a lot, but the masks became
excessively large and I still would lose a few voxels within the brain.
The old
version of FSL (3.3 and before) didn't have this problem. I've tested
this with
data from different scanners and sequences, and I don't think it's simply
because of data quality issues. I suspect anyone using the default
%brain/background in the new version will have the same problem of bad masks
that get worse as you add more subjects.
Here's my earlier thread:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0711&L=FSL&D=0&I=-3&T=0&P=53766
So, I wanted to get prevent losing voxels within the brain and I wanted
to keep
the mask from being too large (i.e., including a lot of outside-the-brain
voxels). I worried about increasing the the search space for the inferential
stats (like you mentioned). I also worried about registration problems since
the poorly masked data might not give the best registration possible
because of
all the stuff on the outside of the brain.
I got around these problems by
1.) turning off BET and setting %brain/background setting to zero in
pre-processing
2.) creating my own mask based off the mean_func.nii.gz file and applying that
mask to the functional data
My workaround assumes you do pre-processing and model fitting
separately. While
you're in the prestats.feat directory, simply enter the following two
commands:
1.) bet mean_func.nii.gz new_mean_func -f 0.45 -m
2.) fslmaths filtered_func_data.nii.gz -mul new_mean_func_mask.nii.gz
new_filtered_func_data
The new_filtered_func_data.nii.gz should be used during the model
fitting stage
of the analysis. Of course you can adjust the fractional intensity
threshold to
adjust the size of the mask. This has worked very well for me and other
folks in
my group. Just remember to skip pre-processing when you do the model fitting
since all that has been taken care of.
Cheers,
David
Quoting "Joseph T. Devlin" <[log in to unmask]>:
> Hi Steve,
>
> Thanks for the extra info.
>
>> Yes - it's just supposed to be a mostly-unimportant, liberal brain
>> masking. It shouldn't matter if it is over-liberal - I've not seen any
>> cases where that was a problem.
>
> Well, in the data I've been working with, it is over-liberal and just
> provides out-of-the-brain false positives. Presumably it has at
> least a small effect on the inferential stats by inflating the search
> space as well. I may try replacing the command with something less
> liberal and see how it affects the results over a couple of studies...
>
>> If the 10% intensity thresholding is too high (it isn't normally?)
>> then you can just lower that under the 'misc' tab, even down to 0 if
>> necessary. Hopefully the voxels that you're losing are just down to
>> the thresholding.
>
> Hard to answer this one. When I bump the threshold up to 30%, it
> removes most of the material outside of the brain. But it also means
> more missing voxels within the brain...
>
>> It's to exclude voxels which are dodgy to use for exactly this
>> reason ;-) Do you really want to include such voxels who's motion is
>> taking them in-and-out of being 'valid'? You can always lower the
>> threshold though :)
>
> I understand your point here but it has unfortunate consequences at
> the group level where any single voxel which was potentially dodgy in
> a single volume in any subject ends up being excluded from the group.
> Also, some of the voxels in my data that are being excluded are not
> due to motion -- they are deep grey matter where the signal is
> slightly reduced relative to the surface. Within that set,
> occasionally a voxel has an abnormally low signal value -- apparently
> randomly. The same process seems to be occurring in surface voxels
> with occasion outlier values, but the overall higher signal intensity
> in these areas makes this not a problem. Basically, the -Tmin option
> seems overly conservative to me -- I'm missing a non-trivial number
> of voxels in my group analyses and it gets worse the more subjects I
> have.
>
> Obviously this reflects data quality in problematic areas and I'm
> working to track that down, but I'm concerned that given the data I
> have, the procedure may be overly-liberal on the edge of the brain
> and overly conservative inside the brain volume.
>
> Anyway, thanks for all the info -- it's very helpful.
>
> All the best,
> Joe
>
--------------------------------------------------
David V. Smith
Graduate Student, Huettel Lab
Center for Cognitive Neuroscience
Duke University
Durham, NC 27708
www.mind.duke.edu
--------------------------------------------------
|