Dear Hamed and Bruno,
Masking your results like this might reduce the number of voxels in the
search volume but not necessarily reduce the adjusted p-values as the
random field theory includes a penalty for the geometry of the search
space (resel counts are defined for each dimension of the search space),
see, eg, page 13 of this article:
http://www.fil.ion.ucl.ac.uk/~karl/A%20unified%20statistical%20approach%20for%20determining%20significant%20signals.pdf
> Thus there is no advantage to applying (3.1) to highly folded
> surfaces for the purposes of decreasing the volume and thereby
> decreasing the P-value, since this simultaneously increases the sur-
> face area. In such cases it is better to surround the region with a
> smoother envelope, at a small cost in volume, but a greater saving in
> surface area and P-value.
Best regards,
Guillaume.
On 02/12/16 12:23, hamed nili wrote:
> hey Bruno,
>
> thanks for your reply.
> I am thinking of consistency across groups and studies: a single GM mask
> that can be used for full-brain group-tests.
> I did try your approach. I got SPM's tissue probability maps for GM
> (given in TPM.nii) and applied a threshold of .25 to it.
> I get something like this. Does this sound reasonable?
>
> Hamed
>
> Inline image 1
>
> On Fri, Dec 2, 2016 at 6:01 AM, Bruno L. Giordano <[log in to unmask]
> <mailto:[log in to unmask]>> wrote:
>
> Hi Hamed,
>
> for the second level my mask is either where the average MNI
> normalized p(gray)>threshold (0.25 is my default reasonable choice -
> check your masks) or, of I use Dartel, where the average of the
> Dartel-space p(gray) > threshold.
>
> If you are not interested in subcortical structures/cerebellum you
> can also use an atlas to take them out (if your group space is
> Dartel you need to deform the atlas from MNI to your group Dartel
> space).
>
> If you want strict consistency between group and native level masks,
> you can finally threshold at the first level by considering the
> group mask back-transformed to native space. This is also really
> handy if you want to remove a given structure from your analysis,
> but don't require the precision you would have when labeling
> uninteresting structures by hand on each subject.
>
> Best,
>
> Bruno
>
>
> On 1 Dec 2016 21:16, "hamed nili" <[log in to unmask]
> <mailto:[log in to unmask]>> wrote:
>
> Dear SPM users/experts,
>
> I have been thinking about this for a while and couldn't come up
> with a definitive answer:
> 2nd level analysis needs to be done in a common space like the
> MNI space. Often there are many voxels that are not interesting
> to us and we are not willing to perform a test on them. This
> means that the MCP could be relaxed by excluding those. These
> include CSF, White matter, etc.
> Now I am aware that at the single-subject level one can take the
> segmentation outputs and apply a mask to the computed results.
> That would automatically take care of that for a single subject.
> My question is about the second-level where we have many
> subjects (say 20). Is there a grey matter template that I could
> constrain my analysis to it? Is there a more elegant way of
> dealing with this?
>
> suggestions are welcome
> Hamed
>
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|