> I just have one question - why might there be large differences in the
> results between the original and optimised VBM techniques?
How different were the un-thresholded t statistic images? They probably
both showed the same general pattern. Also take a look at the beta images
and residuals. The residuals should be slightly smaller for the optimised VBM,
but not uniformly smaller. There may be some areas where the residuals are
higher, but this depends on what you used as a template. In regions with
higher residual variance, you would expect to see lower t statistics and less
sensitivity.
Did you include the volume preservation step (modulation) for the optimised
VBM analysis? If so, then the different results may well relate to different things
being tested. With modulation, the tests are sensitised to absolute differences,
whereas without modulation, the tests are sensitised to the more vague concept
of volumetric differences relative to the overall size of the surrounding region.
Things like ventricle size can also make a difference. With subjects that have
large ventricles, the whole region around the ventricles need to shrink in order
to reduce the ventricular volume. This is because the warps are only parameterised
by about 1000 basis function coefficients. The spatial normalisation in the
"optimised" VBM is based purely on grey matter, so differences purely in ventricular
volume should not influence the warping, whereas when warping whole brains, the
high contrast between white matter and CSF means that the algorithm tries to
shrink this region as much as possible. If there are systematic differences in
ventricular volume, then you may see artificial systematic differences in the volumes
of internal structures that need to grow and shrink with the ventricles. See:
J. Ashburner and K. J. Friston
"Why Voxel-Based Morphometry Should Be Used"
NeuroImage 14(6):1238-1243, 2001.
Maybe also take a look at the Booksteen article in the same issue. VBM is sensitive to
systematic mis-registration as well as volumetric differences. Systematic mis-registration
may differ slightly between the two types of pre-processing. More smoothing tends to
reduce the effects of this though.
>
> In particular, in this months Neuroimage, I have shown that hippocampal
> GMC reduction can be identified using original VBM in patients with
> quantitative evidence of hippocampal atrophy. There was also highly
> significant prefrontal GMC reduction. However, I have re-ran the
> analyses using the optimised VBM technique, and I have lost the
> prefrontal GMC reduction and nearly all of the hippocampal GMC
> reduction. This was done with the same subjects / smoothing kernel size
> etc. I have also gained new 'abnormalities'.
>
> Would it be realistic to maintain that the optimise results better
> reflect actual neuropathology due to the improved normalisation and
> segmentation/extraction?
I would expect spatial normalisation to be more accurate with the optimised version,
and therefore have more faith in the results.
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
|