JISCMail - SPM Archives

RE: spatial normalization and segmentation

Dear Keith,

I grow less confident of this as time goes by, but anyway. I think your post really relates to a question of binary segmentation. If the segmentation algorithm worked in such a way that it simply classified as a voxel as being grey matter (1) or not (0), then your point would be correct. However my understanding is that this is not how it works. Voxels are classified during segmentation with a value in the range 0-1, with voxels that are clearly grey matter generally having a value close to 1, and other voxels having a value close to 0. This value relates to the probability of that particular voxel being grey matter. Therefore virtually all voxels in your segmented image have some kind of value (you can check this by displaying a segmented image and then clicking at various points in the brain). When you come to the analysis stage, you obviously want to exclude from your analysis those voxels with a very low probability of being grey matter. This is what setting the grey matter threshold does.

I've just seen your next posting. My grasp of stats is even worse than my grasp of the rest of SPM, but while I'm here.....the cluster statistic is not considered to be valid for morphometry studies (see Ashburner & Friston, Neuroimage, recently), only the voxel statistics. Additionally, if you have an a priori hypothesis about where you expect to find anatomical differences, it has been advocated that you should use the uncorrected p value in those regions. Alternatively you could use a select a region of a priori interest e.g. the head of the hippocampus, and then use a SVC to test for differences in that structure, which will then give you a corrected p value based only on the ROI you are testing. I think that approach was recently used in a VBM study (Maguire et al, PNAS).

Xavier

-----Original Message-----
From: [log in to unmask] [mailto:[log in to unmask]]On Behalf Of Harenski, Keith
Sent: 15 September 2000 15:16
To: [log in to unmask] '
Cc: [log in to unmask]
Subject: RE: spatial normalization and segmentation

Hello Xavier -

Thanks for the reply and believe me, the length was appreciated. Your gray matter threshold setting suggestion was right on the money. That completely eliminated all "outside" voxels. My only question regarding the grey threshold is, if after we segment the image into the gray compartment based on the a priori images incorporated into spm99, why would I have to specify a gray threshold since all that should be left in the image is gray matter? Or is this simply a way of assuring that any voxels that were misclassified during the segmentation are not included in the analysis?

Thanks again... Keith

-----Original Message-----
From: [log in to unmask]
To: Harenski, Keith
Cc: [log in to unmask]
Sent: 9/14/00 7:22 PM
Subject: Re: spatial normalization and segmentation

Dear Keith,

John Ashburner will be able to give you a much more comprehensive
answer, but I thought I'd add my bit. Firstly, did you use one of the
spm template brains as your template? Those brains include the skull,
and so it might be that the normalisation would not work particularly
well if you are trying to register skull-stripped images. I believe (you
could check in spm_help) that the normalisation initially uses both the
brain and skull in the early stages, and so if it encounters images
without skull this may cause problems. As a more general point, John has
provided a nifty way using the imcalc button to remove traces of skull
from segmented images. From the render button select the grey and white
matter images and choose to save the extracted brain. Then from imcalc
select the grey matter image and then the extracted brain (which begins
brain*) and evaluate the function i1.*i2

It is of course possible that one or more of your images did not
register and segment, which can have a major effect on your findings,
although it sounds like you checked this.

When you set up the SPM analysis it asks you to select the grey matter
threshold (default 0.8). This is used to identify those voxels which
have a high probability of being grey matter, and therefore should be
included in your analysis. It first of all calculates the overall mean
intensity value and then removes all voxels that fall below 1/8 of that
value (I'm not absolutely sure about this, but someone will correct me
if it's wrong), assuming they represent nonbrain tissue. It then
recalculates the mean of the remaining voxels, and then only includes in
the analysis those voxels that survive the threshold you specified. You
may want to look at the mask.img file in the results directory to check
it. Although you can't "limit" the search space (that I know of), you
could alter the threshold to make the criteria for voxel inclusion more
stringent.

You don't say how far the differences are outside the brain, and whether
this is visual on overlay maps, or in terms of xyz coordinates. One
mistake I made in the past was in specifying header information. If you
give the wrong voxel size, it is possible that the Talairach coordinates
will be completely wrong.

There isn't a way of specifying "plugs" of tissue types. The
segmentation already incorporates prior knowledge of the likelihood of
any particular voxel being grey matter (for example), and this is used
to weight the segmentation. If you are finding some sort of smearing
during segmentation, you could try changing the voxel size of the
written out normalised images (via the defaults button), as you may find
that a 1mm cubic voxel size gives you a cleaner segmentation.

Sorry my bit was so long,

Xavier

>Hello all -
>
>I am running a voxel-based morphompetry analysis. After skull stripping
the
>brains i ran them through the spatial normalization, initially with all
the
>set defaults. Subsequent to normailzation the images were semgented
within
>spm into their respective gray, white, and csf compartments then
smoothed.
>When i ran the acutal analysis, spm showed significant voxel intensity
>differences outside of the head. After inspecting each of the
normalized
>brains for any outlying tissue that may have been left after the skull
>stripping and not finding any, i attempted a strictly linear
normalization.
>The results after running the analysis was the same with the same
voxels
>lying outside the head. Additionally any tweaking of the non-linear
>funcions, either in the amount of functions used or amount of
iterations,
>yielded little change in the overall result. Has anyone else seen this?
If
>so, how did you correct for it? I am wondering if there is a way to
limit
>the search space that spm "looks" for significance between the voxels,
that
>i could exclude the area outside of the head.
>
>Also, is there a way to specify "plugs" of brain tissue to classify
voxel
>intenstiy ranges for gray, white and csf prior to running the
segmentation?
>The segmentation result i get is generally very good with only some
>"smearing" of gray and white in the basal ganglia and cerebellum, that
i
>feel could be alleviated if prior knowledge of the intesities of those
areas
>were incorported ahead of time.
>
>thanks in advance for any input.... keith
>
>-------------------------------------------------
>Keith Harenski
>Neurobiology Lab
>WPIC Rm#986
>