Hi Audrey,
You're correct that my masking tools assume that the input masks are normalised.
Smoothing the (normalised) masks a tiny bit (with the standard
smoothing, rather than the one combined inside the new DARTEL2MNI
stuff) is probably also a perfectly good solution. I guess I'm biased
against the intersection approach, but if the smoothing fills all the
holes without creating too much "bleeding" of the mask into the
background, and the normalisation is precise enough that all the masks
agree very well, then the intersection and my alternatives might be
very similar (in the extreme case of all subjects' masks being
identical, then the logical AND or any other fractional combination or
thresholding of the mean mask, etc. etc. will simply return this
identical mask).
You still have the option of computing the intersection/fraction mask
first and using this (same) mask for all of your first level stats
(since you've normalised them) or use slightly different masks for
each first level, then just the intersection/fraction mask at
group-level. It's hard to say which of these would be best, but I'd
expect the differences to be pretty small for these options after your
smoothing.
Best,
Ged
2009/10/16 Audrey Duarte <[log in to unmask]>:
> Hi Ged, thanks for your comments. I made binary masks for each subject prior
> to stats, normalised them to MNI space, which created the holes. As you say,
> the threshold for creating the masks made no difference to the presence of
> the holes after normalisation. What I ended up doing was smoothing the masks
> with a 1mm kernel after I normalised to MNI space and this resulted in masks
> that neither had holes nor were warped outside the FOV. I then applied these
> masked to my smoothed (8mm) normalised (DARTEL) EPIs to eliminate the extra
> brain voxels in those and then submitted those masked EPIs to first level
> stats. Would your tools allow me to create an intersection mask that is
> based on a fraction of all included masks, as you suggest below, and then
> use this as an explicit mask for stats? If I understand you correctly, the
> tools would be used on masks that were first normalised. Thanks, Audrey
>
> On Wed, Oct 14, 2009 at 4:27 AM, DRC SPM <[log in to unmask]> wrote:
>>
>> Hi Audrey,
>>
>> Further to Christian's reply, you might prefer not to apply the masks
>> to the images, but rather to create an explicit mask from the set of
>> masks, and then use this in the stats. This way you can avoid the
>> requirement for voxels to be present in all masks, instead requiring
>> just a certain percentage, or perhaps re-binarising an average of your
>> masks. I have some simple tools that might help with this, and a
>> related paper on masks excluding atrophied regions in VBM analyses:
>> http://www.cs.ucl.ac.uk/staff/gridgway/masking/
>>
>> I'm a bit confused whether your masks are derived from the
>> DARTEL-normalised EPI data (as suggested by doing first level stats
>> after normalisation) or whether they are derived and then normalised
>> (as suggested by "with smoothing for the EPIs without for mask"). In
>> the first case, if the masks are derived from smoothed data, you can
>> probably just reduce your masking threshold, as Christian suggested,
>> and/or use the above approach. However, if you normalise the masks,
>> then it's possible for holes to appear due to the forward-mapping
>> strategy that DARTEL normalise-to-MNI uses. If these holes have value
>> zero,
>> (as in
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0909&L=SPM&P=R64832
>> point 1)
>> then no amount of lowering the threshold will help, so you might
>> prefer to lower the fraction of masks in which you require the voxels
>> to appear instead, using the above-mentioned tools. This is assuming
>> that the smoothing of the normalised EPIs (or contrast-images)
>> themselves means that these voxels do contain valid data, despite the
>> mask being zero there.
>>
>> All the best,
>> Ged
>>
>>
>> 2009/9/2 Audrey Duarte <[log in to unmask]>:
>> > Hi Christian, I wanted to weigh in with my own results and ask about
>> > your
>> > mask. I followed this procedure you suggest, with a few changes. I
>> > created
>> > binary FOV masks myself for each subject, since I typically do first
>> > level
>> > stats after normalisation. I applied these masks to the EPIs after the
>> > flow
>> > fields had been applied (with smoothing for EPIs without for mask). This
>> > masking did take care of eliminating voxels outside the FOV as does your
>> > procedure. However, after applying the flow fields (DARTEL deform to
>> > MNI) to
>> > the mask without smoothing, little holes appear in the image which
>> > subsequently appear in the EPIs after applying the mask. Did you see the
>> > same in your mask after deforming it to MNI space?
>> >
>> > The images from L to R are the EPI deformed to MNI with smoothing, the
>> > FOV
>> > mask for the same subject, the deformed mask without smoothing, single
>> > subject canonical brain.
>> > Thanks, Audrey
>> >
>> > 2009/9/1 Christian Büchel <[log in to unmask]>
>> >>
>> >> Dear Jonathan and Marko,
>> >>
>> >> The problem seems to stem from the combined warping and smoothing that
>> >> is
>> >> performed. It always occurs when your EPIs are truncated. Doing some
>> >> control
>> >> analyses revealed that inside the original EPI everything is correct.
>> >> It's
>> >> just that the deformation field enlarge the voxels outside the FOV.
>> >> We have designed a pipeline that takes care of these problems (and at
>> >> the
>> >> same time saves a lot of disk space as only the con images are
>> >> normalized
>> >> and smoothed...).
>> >>
>> >> 1. Coregister T1 onto first functional image
>> >>
>> >> 2. Realign and reslice fMRI, then do first level stats (unsmoothed
>> >> images!)
>> >>
>> >> 3. Use all T1 images --> new segment --> Dartel
>> >>
>> >> 4. Take individual con images from first level analyses and apply
>> >> deformations (and smooth) them into MNI space (using "normalise to
>> >> MNI").
>> >> These images will again look funny.
>> >>
>> >> 5. Take the individual mask.img (that was created by the 1st level
>> >> stats)
>> >> and also deform this into MNI (using Dartel) BUT without smoothing.
>> >>
>> >> 6. Take all spatially normalised individual mask.img and do a logical
>> >> AND
>> >> (using IMCALC) with all of them and use the ensuing image as a mask for
>> >> the
>> >> final 2nd level stats.
>> >>
>> >>
>> >> I hope this works for you
>> >>
>> >> -Christian
>> >>
>> >> --
>> >> Prof. Dr. Christian Büchel
>> >> Institut für Systemische Neurowissenschaften
>> >> Haus W34, Universitätsklinikum Hamburg-Eppendorf
>> >> Martinistr. 52, D-20246 Hamburg, Germany
>> >> Tel.: +49-40-7410-54726
>> >> Fax.: +49-40-7410-59955
>> >> [log in to unmask]
>> >>
>> >> http://www.uke.uni-hamburg.de/institute/systemische-neurowissenschaften/
>> >>
>> >> > -----Ursprüngliche Nachricht-----
>> >> > Von: SPM (Statistical Parametric Mapping)
>> >> > [mailto:[log in to unmask]] Im Auftrag von Jonathan Peelle
>> >> > Gesendet: Dienstag, 1. September 2009 14:57
>> >> > An: [log in to unmask]
>> >> > Betreff: [SPM] SPM8: DARTEL normalise to MNI problem with
>> >> > partial coverage EPIs
>> >> >
>> >> > I've come across a curious normalization issue using the
>> >> > "normalise to MNI" option in SPM8 (with recent updates) on
>> >> > partial-coverage EPI scans. Segmentation, coregistration,
>> >> > and template creation seem to work ok (see struct_c1_exampfun.png).
>> >> >
>> >> > Applying the flow field to the segmented structural image
>> >> > seems to be fine. However, when I apply the flow field to
>> >> > the functional data, I get the strange warping attached (see
>> >> > avg152_mwc1_swf.png). This came up at least once previously
>> >> > on the list, but I didn't see any replies:
>> >> >
>> >> > https://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0908&L=SPM&P=R
>> >> > 399&X=485D484296C56BEE64
>> >> >
>> >> > Have others run into this issue? Any idea what this might be
>> >> > due to, and (even better) possible suggestions for ways to
>> >> > fix? Thanks!
>> >> >
>> >> > Jonathan
>> >> >
>> >> > --
>> >> > Jonathan Peelle, PhD
>> >> > MRC Cognition and Brain Sciences Unit
>> >> > 15 Chaucer Road
>> >> > Cambridge CB2 7EF
>> >> > UK
>> >> >
>> >>
>> >>
>> >>
>> >> --
>> >> Pflichtangaben gemäß Gesetz über elektronische Handelsregister und
>> >> Genossenschaftsregister sowie das Unternehmensregister (EHUG):
>> >>
>> >> Universitätsklinikum Hamburg-Eppendorf
>> >> Körperschaft des öffentlichen Rechts
>> >> Gerichtsstand: Hamburg
>> >>
>> >> Vorstandsmitglieder:
>> >> Prof. Dr. Jörg F. Debatin (Vorsitzender)
>> >> Dr. Alexander Kirstein
>> >> Ricarda Klein
>> >> Prof. Dr. Dr. Uwe Koch-Gromus
>> >
>> >
>> >
>> > --
>> > Audrey Duarte, PhD
>> > Assistant Professor
>> > School of Psychology
>> > Georgia Institute of Technology
>> > 654 Cherry Street
>> > Atlanta, GA 30332
>> > voice 404-894-2349
>> > http://psychology.gatech.edu/duartelab/
>> >
>
>
>
> --
> Audrey Duarte, PhD
> Assistant Professor
> School of Psychology
> Georgia Institute of Technology
> 654 Cherry Street
> Atlanta, GA 30332
> voice 404-894-2349
> http://psychology.gatech.edu/duartelab/
>
|