I have not followed the mail exchange on your question so far, so
hopefully this is not redundant.
As you write, you can extract 'raw data' from the smoothed images by the
V. O. I. function, e. g. for one voxel (radius 0 mm) or some other shape.
The meaning of these values depends of your preprocessing: I assume, your
data have undergone the Jacobian modulation step that is generally
recommended for VBM - multiple postings exist on this issue. If you had
switched on this step during the normalisation or segmentation step, the
values extracted can be interpreted as volumes and can therefore be
treated like that in an offline statistics.
Depending on what your question is, you would then probably also have to
include appropriate covariates, including total grey (or white) matter,
gender, age or others.
If you had not switched on the modulation step and selected 'preserve
concentration' instead of 'preserve volume', the interpretation of the
local value is indeed a different one, rather a probability of that voxel
to be GM.
Does this answer the question?