> Meanwhile, are you sure of the below?
>
> > > I am analyzing a group of SPECT scans consisting of about 50 scans. The
> > > problem is that a portion of these scans (14 of them) are slightly to
> > > mildly cropped (i.e. part of the cerebellum is cut off). I am
> > > wondering how this will effect my results and/or what I can do about
> > > this problem. Based on previous emails, it looks like one suggestion
> > > might be to create an object mask, if that is a possible solution how
> > > do I go about creating one?
> >
> > It shouldn't have too much of an effect, other than the region of brain
> > included in the final analysis will be smaller. I would only suggest any
> > kind of "object masking" if there is a region within the image volume
> > that is not how it should be. I am assuming that the part of the
> > cerebellum that is cropped is outside the image volume.
>
> Lets say you have n subjects in group A and m subjects in group B. Also say
> that there has been a systematic difference in axial positioning such that
> all subjects in group A has been scanned 1cm (2 slices) more cranially than
> group B. After spatial normalisation this means there will be a few planes
> containing non-zero values in group A, but with zero-values in group B (or
> is there some way to apply masking in the spatial normalisation (similar to
> the realignment masking, but across subjects rather than scans) ?).
> If we did the statistics there and then it would not be a problem since the
> implicit intensity based mask would remove all those voxels. But we don't.
> Instead we spatially smooth the data. If we consider a voxel at the very
> edge of the area where we have data for both subjects then we will have a
> different mixture of brain and non-brain values within the filter-kernel
> for the two groups. Also, especially for the "inner part of the edge" where
> we have all brain values for group A and only a few non-brain values for
> group B within the kernel it is likely that voxels will pass the conditions
> for the implicit mask. Statistics on this will now compare apples and
> pears.
>
> Knowing you I know you will have thought of this, and I am sure there is
> some clever way in SPM to avoid this problem. However, I haven't found it
> and I suspect perhaps some others haven't either. I have actually helped
> some people (in Uppsala and Huddinge who did SPECT and FDG PET respectively
> on different dementia groups) with this very problem. When they just
> performed a "straightforward" analysis there were quite massive differences
> at the edge of the "actual axial FOV" of the kind one would expect from the
> reasoning above. By AND-masking all the images prior to spatial smoothing
> these differences dissapeared.
> Have I been doing something stupid (again) to "fix" something that SPM
> could have handled?
Good point. I forgot about this (got mixed up by object masking in spatial
normalisation). The data probably should be masked between the spatially
normalised data being written, and it being smoothed. The spm_mask.m
routine can be used to do this:
P1 = spm_get(Inf,'*.img','Images to define mask from');
P2 = spm_get(Inf,'*.img','Images to mask');
spm_mask(P1,P2);
The 'Images to define mask from' would be an example spatially normalised image
from each subject, whereas 'Images to mask' would be all the images to enter into
the analysis. The output masked images are prefixed by 'm'.
Best regards,
-John
--
Dr John Ashburner.
Functional Imaging Lab., 12 Queen Square, London WC1N 3BG, UK.
tel: +44 (0)20 78337491 or +44 (0)20 78373611 x4381
fax: +44 (0)20 78131420 http://www.fil.ion.ucl.ac.uk/~john
|