> I am trying to conduct a comparison between patients and controls using VBM
> on DTI data. I have a single FA image for each subject that only includes
> brain matter, and everything outside the brain is masked out.. What would
> be the best way to process this data for VBM? Since this data is already
> skull-stripped, I assume that I will not need to segment and brain extract
> before normalization.
If you plan to use VBM, why not use T1 weighted images, rather than FA maps?
An SPM analysis of FA maps is not really VBM as I understand it, as these
data don't reflect the volume of some tissue type. The idea of VBM is that
data are pre-processed in such a way that subsequent statistical tests are
sensitised to volumetric differences.
>
> The sequence of steps I had in mind is as follows:
> 1. Coregister FA maps in AIR
What are you coregistering the FA maps with?
> 2. Normalize to EPI template in SPM
I don't know how similar the FA maps are to the EPI template.
This may work, or it may not.
> 3. Create tissue segments... I'm not sure if this segmentation routine is
> dependent on intensity or priors. If someone could elaborate on this
> further, that would be great. Also, it would make sense not to correct for
> intensity inhomogeneities at this step, but I cannot be sure.
The segmentation in SPM is based on voxel intensities and on the priors.
Intensities of different tissue types are supposed to be approximately
Gaussian. The algorithm estimates the means and variances of these
distributions for the various classes. From these (making use of the
priors), it assigns probabilities of the voxels belonging to each class.
As an example, the probability of a voxel being grey matter given that
it has an intensity of "y" is:
P(gm|y) = P(y|gm)*P(gm) /
(P(y|gm)*P(gm) + P(y|wm)*P(wm) ... + P(y|o3)*P(o3))
P(y|gm) means the probability of y given that the voxel is in the grey
matter class. This is based on:
P(y|gm) = exp(-(y-m)^2 /(2*v))/sqrt(2*pi*v)
where "m" is the mean and "v" is the variance of the GM class.
P(gm) is based on the prior belonging probability for GM, and also on the
proportion of voxels in the GM class.
I suspect that the Gaussian distributions don't work so well for FA maps,
resulting in not-so-good segmentations.
> 4. Smooth white matter segments
> 5. Use smoothed WM segments for statistical comparison.
>
> I'm not sure if this is the correct way to perform this analysis, and I'm
> not sure how John's scripts would react to FA maps as opposed to T1
> anatomicals.
The script probably wouldn't work so well.
Best regards,
-John
--
Dr John Ashburner.
Functional Imaging Lab., 12 Queen Square, London WC1N 3BG, UK.
tel: +44 (0)20 78337491 or +44 (0)20 78373611 x4381
fax: +44 (0)20 78131420 http://www.fil.ion.ucl.ac.uk/~john
|