Dear All,
> I’m not convinced that this outcome is primarily due to the mismatch
> between the characteristics of your study population and the template
> (although it might be aggravated by it):
I fully agree. Your imanges look like they were not normalized correctly
(remember that even for native-space segmentation, the priors are in
efect inversely normalized to the input images). Your results look like
they are mainly the result of the prior probability maps. This can
happen when matching the priors goes completely wrong which again can
happen when you have very inhomogenous input data. This is not
high-field data by any chance?
> Is it possible that you didn’t set the origin before preprocessing the
> data? In our experience the unified segmentation approach is pretty
> vulnerable to end up with such deformed results if you don’t give the
> program a reasonable starting point. Assuming that the data quality is
> o.k., my best guess would be: reset the origin to AC, rerun data
> preprocessing for these subjects, and you’re done.
This may help, too. Christian had implemented an automated determination
of the center of mass of an image in order to improve the starting
estimates in a beta-version of the 5.1 toolbox but I don't know if it is
out yet.
Also,
> The ages of the subjects are 1-3 years old.
there is no reference data for such subjects yet that I am aware of. We
(Christian, the Cincinnati IRC group and myself) have two abstracts at
HBM where we investigate a prior-less segmentation for datasets from
infants. In effect, the segmentation is done as in spm5 but for writing
out the results only tissue intensity information is used, disregarding
the influence from the priors. It does make the segmentation somewhat
more vulnerable, especially w.r.t. inhomogeneity, but it seems to
considerably improve segmentation results on "unusual" datasets. Not
sure if the updated 5.1 toolbox is already publicly available but you
could ask Christian.
> The segmented images come out like this (please see attached jpg of
> brains with modulation, segmentation and normalization applied; the left
> brain is the one with the problem, the right brain is for comparison).
Ah, good to know ;) But again, this is a problem with matching the
priors and thus effectively a normalization problem. Note the oblique
cutoff at the upper left side, this is where either the FOV or the
tissue boundary of the input image went. Try the suggestions above and
let us know how you fared.
Best,
Marko
--
=====================================================================
Marko Wilke (Dr.med./M.D.)
[log in to unmask]
Universitäts-Kinderklinik University Children's Hospital
Abt. III (Neuropädiatrie) Dept. III (Pediatric neurology)
Hoppe-Seyler-Str. 1, D - 72076 Tübingen
Tel.: (+49) 07071 29-83416 Fax: (+49) 07071 29-5473
=====================================================================
|