FYI: The FreeSurfer folks have also found that the issue of bias in
longitudinal analysis due to asymmetric treatment of scans is an
appreciable problem. Thus, the most recent versions of the longitudinal
analysis steam in FS are designed to be unbiased with respect to any
time point.
http://surfer.nmr.mgh.harvard.edu/fswiki/LongitudinalProcessing
cheers,
-MH
On Wed, 2010-10-13 at 16:31 +0100, John Ashburner wrote:
> DARTEL is not so suited to longitudinal registration as it works at a
> lower resolution than HDW and because it is based on aligning tissue
> classes. Both of these mean that it may not be as accurate at
> estimating the very small deformations that occur within subject over
> time.
>
> There are many many ways to do a longitudinal analysis, but here is
> one approach that may be acceptable, which involves identifying
> regions of longitudinal change in GM.
>
> 1) Coregister the early and late scans of each subject together. This
> will provide the initial rigid body alignment that HDW will use.
>
> 2) Run the HDW within subject. This could involve registering the
> early to the late scan, or the late to the early scan. Because this
> registration is not exactly inverse consistent, this choice is likely
> to change the findings slightly. Lets say that the late scan is
> registered to the early one (ie early image stays fixed and late image
> is warped to match it). This will generate a map of Jacobian
> determinants (j*.img) that encode relative volume changes. Larger
> determinants (greater than one) will encode regions of growth. Values
> less than one encode regions of shrinkage.
>
> 3) Segment the early scan to generate grey and white matter, as well
> as "imported" grey and white matter, which will be used by dartel.
>
> 4) For each subject, create a map of GM volumetric difference. This
> can be done using ImCalc and involves subtracting the grey matter from
> the early scan from the amount of grey matter that we would expect
> from the late scan. The early time point GM is simply what is in the
> c1 image. Assuming accurate segmentation and longitudinal
> registration, the grey matter in the late time point can be computed
> by multiplying the Jacobian determinants by the c1 image. Putting
> this all together, you would select the j image and the c1 image, and
> evaluate
> i2.*(i1-1)
>
> Alternatively, you may wish to just use the volumetric difference,
> which would be by selecting the Jacobain image and evaluating
> (i1-1)
>
> If the time difference between the scans is variable, then you could
> also normalise these differences by dividing by the time between the
> scans. This may simplify the design matrix when you fit the GLM,
> although it does represent a slightly different model.
>
> 5) After all the within subject preprocessing is done, you can dartel
> all the early data together (ie run dartel to align all c1 scans to
> the group average GM, while simultaneously aligning the c2 to the
> group average WM).
>
> 6) Use the normalise to MNI space option of dartel to generate
> smoothed Jacobian scaled spatially normalised versions of the images
> generated in (4).
>
> 7) Do the stats.
>
> Note the asymmetric treatment of the early and late scans. To be a
> bit more rigorous, you could repeat the analysis with early and late
> scans swapped around to see if you get the same results. See the
> following paper for more info....
>
> Bias in estimation of hippocampal atrophy using deformation-based
> morphometry arises from asymmetric global normalization: An
> illustration in ADNI 3 T MRI data
> NeuroImage, Volume 50, Issue 2, 1 April 2010, Pages 434-445
> Paul A. Yushkevich, Brian B. Avants, Sandhitsu R. Das, John Pluta,
> Murat Altinay, Caryne Craige and the Alzheimer's Disease Neuroimaging
> Initiative
>
> Another approach would be to pre-process the data both ways and create
> weighted averages of the smoothed spatially normalised Jacobian scaled
> difference images (depending how the Imcalc bit is done the second
> time, you may need to subtract one from the other in the averaging
> procedure). Then do the stats on these averages. This should (I
> hope) solve the problem reported by Yushkevich et al, and make better
> use of the data.
>
>
> Best regards,
> -John
>
> On 7 October 2010 09:46, Emma Burton <[log in to unmask]> wrote:
> > Dear John and SPMers
> >
> > I read with great interest the previous posts on using DARTEL in SPM8 for longitudinal VBM. However I am a little confused as to which is the best way of processing the data. Should DARTEL be used for this? and how does the modified version of the HDW approach that Kipps used get incorporated into DARTEL.
> >
> > I am new to DARTEL and HDW and wondered if any one could help with details of the steps needed to perform longitudinal VBM.
> >
> > Any help would be appreciated. Many thanks in advance.
> >
> > Emma
> >
|