> I would like to use DARTEL to normalize fMRI-data of my patients.
> We had some discussion about this in our group; therefore, I would be pleased if you could help me with my questions below.
>
> 1) We are working with multiple sclerosis patients who have multiple structural small lesions scattered over the entire brain. I assume that normalization methods suffer in their accuracy when applied to neurological patients anyway, but I expected DARTEL to be the better choice compared to the default spatial normalization strategy in spm8. However, a colleague of mine suspected that DARTEL would introduce MORE errors because the template generated could be somehow distorted due to the various lesion locations. Do you agree? And what would you say about applying DARTEL to patients with larger lesions or neurological patients in general?
The main thing to watch out for is the segmentation part, before you
actually run Dartel. Providing this is able to ignore the lesions,
and include them in the WM component, then all should be reasonably OK
(or at least better than it would be otherwise).
In general though, if there are systematic differences between the
groups in some part of the brain, then the pre-processing is unlikely
to fully correct this, leading to differences in other parts of the
brain also being detected within the mass-univariate analysis
framework. As far as I can tell, the only way to circumvent this is
to fit a proper generative model of the MRI scans. This would allow
the registration and segmentation components of the model to properly
account for the confounding effects of the certain differences between
the groups.
In practice, we don't do this. Instead we usually apply a pipeline of
procedures to the data, such that the output of each step serve as
features that are input to the next step. This means that differences
can theoretically show up in the "wrong place" after mass-univariate
testing. There is evidence for this sort of effect in papers such as
this one:
Freire L, Mangin JF (2001): Motion correction algorithms may create
spurious brain activations in the absence of subject motion.
Neuroimage 14:709–722.
A partial solution to the problem is described here:
Freire, Luis, Jeff Orchard, Mark Jenkinson, and Jean-François Mangin.
"Reducing Activation-Related Bias in FMRI Registration." In Medical
Imaging and Augmented Reality, pp. 278-285. Springer Berlin HeidelIn
general, if there are systematic differences between the groups in
some part of the brain, then the pre-processing is unlikely to fully
correct this, leading to differences in other parts of the brain also
being detected within the mass-univariate analysis framework. As far
as I can tell, the only way to circumvent this is to fit a proper
generative model of the MRI scans. This would allow the registration
and segmentation components of the model to properly account for the
confounding effects of the certain differences between the
groups.berg, 2004.
> 2) If I used DARTEL nonetheless, I assume that I would have to separate patients and healthy controls and generate a template for each group independently. Is that correct?
No! Within the pipeline approach, for any subsequent statistical
tests to be valid, you must have treated all data in an identical way.
If you don't do this, then the pre-processing will reduce the
residual variance within group, while not reducing the variance
between group.
> 3) Data is acquired successively over several years. To apply DARTEL I probably have to wait until I get the data of the last subject or is there a way to solve this issue differently, e.g. by using an existing template?
There are existing templates you could use, but note that they may be
systematically different from your data. Slight differences in MR
contrast will change the behaviour of the tissue segmentation. For
example, the tissue priors will have more influence on the final
segmentation if images are more noisy (as the intensities are deemed
less reliable indicators of tissue class).
> If so, is there an existing representative template available anywhere?
It really does not take that long to run Dartel. I had to run it
hundreds of times when debugging the code.
> And in that case, would it still make sense to use DARTEL or should I stick with the default strategy?
If you have longitudinal data, you may want to take a look at the new
toolbox in SPM12b.
Best regards,
-John
|