Hello all,
I don't know how often somebody replies to a message 2 years later, but
here goes. Has everyone just been using the default bounding box when doing
spatial normalization? It seems a bit strange to me that the standard would
be to cut off and "ignore" activation at the bottom of the cerebellum. I
hadn't known that SPM was doing this (I should have looked more carefully
when I checked the registration of the images), until I saw a line of
activation in one subject's contrast map at the border of the bounding box
in the cerebellum - activation that was (probably) part of a bigger cluster
outside the bounding box. The inferior portion of the cerebellum has been
implicated as being involved in verbal working memory, amongst other things
(I know John Desmond has a few papers that highlight this region), so I
think it's probably important to a lot of the imaging community, if not all.
I'm using SPM2 (so I don't know what SPM5 is doing), and I have chosen to
process all of my data instead using the SPM "template" bounding box, which
encompasses the whole template. I feel like that should be the "default"
rather than a bounding box which clips the cerebellum - does anybody have
any opinions on this? Thanks!
Daniel
Daniel Simmonds
Developmental Cognitive Neurology
Kennedy Krieger Institute
[log in to unmask]
On Fri, 2 Jul 2004 10:46:35 +0200, Jesper Andersson <[log in to unmask]>
wrote:
>Dear Kosha,
>
>>I did my analysis using SPM2 processing. Images were of the dimension:
>>40X64X16 and voxel size of 3.75X3.75X5. Spatial Normalisation using the
>>default bounding box gave me the output image as 79X95X69 and voxel size
>>2X2X2. I ran the statistics and got the individual results. I did group
>>analysis on 14 subjects. I however lost the fusiform activation i was
>>getting in the group results comparing them to the individuals. I
>>increased the Bounding Box of normlisation step to the same as template,
>>meaning that my output image was now 91X109X91 and voxel size of 2X2X2.
>>Doing the statistical analysis followed by the grp gave me back the
>>fusiform activation i missed in the previous analysis. is there any
>>explanation to this?? My smoothing was fwhm=8mm.
>>
>>
>>
>The matrix size of the template is quite "generous" in that it contains
>a quite a bit of air surrounding the "standard brain". This is *not*a
>problem in itself, and the main reason that the default bounding box is
>smaller is to save disc space.
>
>The default bounding box is big enough to accomodate the whole brain,
>save for the very basalmost bits of the cerebellum. It is definately big
>enough to contain the fusiform. I hardly think the explanation for your
>problem is that any parts of your activation has been chopped off.
>
>Depending on how the bounds of your bounding box was defined the new
>voxel-centres may or may not coincide for the two different boxes. If
>they dont the resampled values will be slightly different for the two
>cases and if your activation goes from "just super-threshold" to "just
>sub-threshold", then that is a concievable explanantion.
>
>>I think that the bounding box is maing a lot of difference as it clips off
>>certain part of the brain you might be interested in. so choosing the
>>default might not always be a good idea. But then does choosing a bigger
>>bounding box affect any other way the analysis??
>>
>>
>>
>As I said, if everything works the default bounding box should clip only
>a tiny part of the cerebellum. Choosing a bigger box might increase the
>search volume by a tiny amount, but other than that it should have no
>adverse effects on your analysis.
>
>Just one extra thing.
>
>Between SPM99 and SPM2 the defaults for "masking" in the spatial
>normalisation changed. In SPM99 a mask was used to ensure that only
>brain-voxels were considered when calculating teh sum of squared
>differences between the template and the "object" brain. In SPM2 it is
>not (by default).
>
>That means that if you use a structural (e.g. T1 weighted) to estimate
>spatial normalisation parameters in SPM2 the results may not be perfect.
>The reason for that is that normalisation is a compromise when it comes
>to aligning different structures (because the displacement field has a
>limited spatial resolution).
>
>So, consider the SPM template which has a reasonably sized brain and a
>typical skull thickness. Then imagine we want to normalise a subject
>with a tiny brain and a really thick skull (not unusual combination).
>SPM would then like to expand the object image (to increase his brain
>size), but at the same time SPM would like to shrink parts of the object
>image (to match the skull thickness to the template). The end result is
>a compromise where neither the brain nor skull are ideally matched.
>
>This is *not* a problem when using your EPI images to normalise to the
>EPI template (cause non-brain is more or less invisible and doesn't
>affect the process).
>
>For those using the structurals for the normalisation I would recommend
>comparing the results when using the default (no masking) and when
>changing the default, by editing spm_defaults.m so that it reads
>
>defaults.normalise.estimate.weight = '"whatever"/apriori/brainmask.mnc';
>
>When I have seen "clipping" it has often been related to the
>normalisation going a little wrong (e.g. for the reason suggested above).
>
>Good luck Jesper
>=========================================================================
|