The template was generated by:
1) Segment and affine register some CT data from various sources with
the MNI-space.
2) Combine the GM, WM and CSF of SPM5, with the soft tissue and bone
data from the CT data. This formed the first iteration of the new
version.
3) Segment the T1w scans (same subjects as in the Dartel paper, and
mostly the same as used in Good's paper) using the approximate priors.
4) Dartel everything together.
5) Manually edit the sufficient statistic images (using MRIcron),
setting voxels to zero where the values were clearly misclassified.
6) Used the smoothing method described in the "Creating average shaped
template" paper to generate crispish tissue probability maps from these.
7) Projected the crisper tissue probability maps over the individual
scans using the inverse Dartel transform, using them to re-segment the
original scans.
8) Affine transformed all the new tissue class images to MNI space.
9) Computed sufficient statistics, and applied the smoothing from the
"Creating average shaped template" paper, but with a bit more
regularisation (appropriate for affine registered scans).
The whole thing took about a week, and it's probably not really
publishable. The scans had fairly low grey/white contrast, so the
resulting tissue probability maps may not be as optimal as they could
be. I hope to repeat the whole procedure using some higher quality data
and a bit more mathematical rigour in the algorithm.
Best regards,
-John
On Thu, 2010-02-18 at 10:38 +0100, Ribes Delphine wrote:
> Dear SPM_user, John Ashburner,
>
> I've been using the New semgent options in SPM8 and I have some questions regarding the probabilistic template.
>
> In the article Unified segmentation from John Ashburner 2005, which explain how new segment works, I've read :
>
> The data we
> used represent a modified version of the ICBM Tissue Probabilistic
> Atlas.2 They consist of tissue probability maps of grey and white
> matter and of CSF (see Fig. 1). A fourth class is also used, which is
> simply one minus the sum of the first three. These maps give the
> prior probability of any voxel in a registered image being of any of
> the tissue classes—irrespective of its intensity. The current
> implementation uses tissue probability maps for grey matter, white
> matter, and CSF, although maps for additional tissue types (e.g.,
> blood vessels) could also be included. The simple model of grey
> matter being all of approximately the same intensity could also be
> refined by using tissue probability maps for various internal grey
> matter structures (Fischl et al., 2002).
>
> In the SPM8 new segment toolbox, I've seen that it is using 6 probabilistic maps: GM, WM, CSF, Other1, Other2 and Background.
> For those template, did you also used a modified version of the ICBM Tissue Probabilistic Atlas? Did you use additionnal information? Did you use additionnal atlas?
>
> I would like to publish some results which use New Segment from SPM8 and so I would like to have the latest update concerning the template used!
>
> Thanks a lot,
> Have a good day,
> Delphine
--
John Ashburner <[log in to unmask]>
|