> I used the "Apply deformations" from the "Extended tools" option in the VBM8
> toolbox to map a contrast obtained with VBM into the native space of a
> subject again. I used the iy_r*.nii deformation image but the output image
> (the contrast in the native space) appears with a lot of small blobs around
> the bigger clusters... why does this happen?

I'm not sure, but you may want to try applying the inverse transform
to some other spatially normalised image to make sure things are
working correctly.

>
> I would also like to know if there is a pipeline or a DBM toolbox? I didn't
> find it on the SPM archives and I really don't understand how should I
> perform the DBM analysis (inputs, outputs, parameters, etc.)

Technology has moved on a bit since that old DBM paper.  Now, I would
suggest doing it using some form of whole brain multivariate analysis
of the velocity fields generated by Dartel (or, better, with my new
algorithm if the paper ever gets accepted).  If you run Dartel, it
will generate a bunch of flow fields (u_*.nii).  The idea would be to
do a multivariate analysis of these.  Among the Dartel tools is an
option to generate a kernel matrix from these flow fields.  Run this
and it will generate a MATLAB .mat file that contains an NxN matrix,
where N is the number of subjects.  This can be decomposed into a
number of parameters for each subject using svd, and then, for
example, you could use the first few columns of this as input to a
canonical correlation analysis.  Unfortunately, there are no simple
ways (that I'm aware of) of doing this via the SPM user interface.

There are a variety of approaches to doing multivariate analysis,
which may be broadly divided into generative and discriminative
models.  If the question is simple (of a sort that can be expressed as
a t-test in a univariate case), then you can just use a simple
multivariate pattern recognition technique, which may be thought of as
a discriminative approach.  One such example is in Karl's MVB paper:

"Bayesian decoding of brain images"
NeuroImage, Volume 39, Issue 1, 1 January 2008, Pages 181-205
Karl Friston, Carlton Chu, Janaina Mourão-Miranda, Oliver Hulme,
Geraint Rees, Will Penny, John Ashburner

For even simpler questions concerning comparisons between two groups
of subjects (that don't consider any confounding effects), there are
approaches such as SVMs.  The idea here is that the algorithm may be
trained on one set of data so that it learns a hypothesis about the
difference between the groups.  This hypothesis may then be tested on
another set of data, and a p value determined.  If data is limited,
then an approach such as leave-one-out cross validation may be used.

Another approach would be to use Bayesian pattern recognition
approaches (such as MVB).  One framework that works well in many
situations is that of Gaussian Process Models (
http://www.gaussianprocess.org/gpml/ ).  These give a measure of model
evidence that may be used for Bayesian model comparisons.

Eventually, this kind of thing will become a bit more streamlined
within SPM, providing I can figure out a good way of visualising
multivariate patterns of shape difference.

This paper may give some more ideas:

"Multivariate models of inter-subject anatomical variability"
NeuroImage, In Press, Corrected Proof, Available online 27 March 2010
John Ashburner, Stefan Klöppel

Best regards,
-John