> I have a 'stupid' question: for how many millimeters could we say the
> SPM5 segmentation can introduce errors (any ref?)
This is entirely data dependent. Statistics are presented for some simulated
data in the "Unified Segmentation" paper, but the behaviour with real data is
likely to be different depending on signal/noise, artifacts etc.
The current "gold standard" for assessing image segmentation algorithms
involves comparing their results against manually segmented images. This
should imply that human judgement is perfectly acceptable for assessing how
well an image is segmented.
It also depends on what is meant by the term "grey matter". The segmentation
algorithm defines it in terms of a mathematical model of the data. The only
information it has about grey matter is in terms of the tissue probability
maps in the spm5/tpm directory.
> I have a reviewer saying that effects observed in the visual cortex and
> cerebellum in a VBM study of mine can be attributed to segmentation
> error (well obviously it would also mean that the error was always in
> the same direction ... )
The reviewer has a perfectly plausible explanation. The differences could
also be attributed to registration errors (see Bookstein's criticism of VBM),
cortical thickness differences, folding differences, differences in the MR
properties of the tissue, partial volume effects etc. A significant
difference among the pre-processed data has many possible explanations, but
it would essentially predict where the null hypothesis (of no difference)
would be deemed unlikely to hold if the same pre-processing was applied to
similar images of a different set of subjects from the same populations.
For VBM studies, such differences are usually explained in terms of tissue
volume, but other explanations are always possible. Similar issues apply in
all branches of science.
Best regards,
-John
|