> I have noted all of your names in prior discussions on the SPM list
> server about normalizing patient data to a stereotactic template, but I
> haven't found much since some time ago. I intend on collecting blood flow
> PETs of healthy normal elderly and AD patients (mild, early stage) and
> would like to compare activation in one group to another. I will have SPGR
> MRIs on the subjects, as well. Can someone tell me the latest thinking on
> the stereotactic normalization of patients with atrophy and elderly, and
> the potentially biasing effects of atrophy using SPM99?
>
> Should I be creating my own template? I have seen reference to
> co-registering the MRI with the PET first, normalizing the MRI, and then
> applying the same normalization paramters to the PET. Is there a preferred
> way to go? Is there literature on this?
Spatial normalisation in SPM99 only involves parameterising the warps with 1176 (which is 3 times
7x8x7) basis function coefficients, which isn't enough to obtain really precise registration.
High frequency warps are not modelled, so when spatially normalising T1 weighted MRI (which includes
a lot of signal from outside the brain) the variability of skull thickness and other non-brain factors can
influence how well SPM's spatial normalisation registers the brain. For this reason, it is often better to
base the estimate of spatial normalisation parameters from an image that has non-brain tissue removed,
by matching it to a template consisting of only brain signal. This is the approach that Catriona Good adopted
for her "Optimized VBM protocol" in:
A Voxel-Based Morphometric Study of Ageing in 465 Normal Adult Human Brains
Catriona D. Good, Ingrid S. Johnsrude, John Ashburner, Richard N. A. Henson,
Karl J. Friston, and Richard S. J. Frackowiak. NeuroImage 14, 21 36 (2001)
This approach involves segmenting grey matter from the images and estimating the warps that best register
grey matter together.
The approach we usually use for spatially normalising PET data is simply to warp the images to an appropriate
PET template. This works reasonably well for us because we have a good field of view, reasonably high signal
to noise and the same contrast in the template as the images we are normalising (because the template was
derived from our own data). As SPM99's spatial normalisation only estimates the overall shape of the brains,
then high frequency signal in the images only confounds the estimation of the warps. We therefore smooth the
data first. This also has the advantages of making the optimisation more efficient, as the Taylor series
approximations become less approximate when the cost function becomes more quadratic. (A more quadratic
cost function also implies an error distribution that is more Gaussian, which means that the MAP estimate from
the spatial normalisation becomes closer to being the expectation of the posterior probability - which is good).
Because we are not using high frequency information from the images, then we don't necessarily need to use
high resolution MRI. You may want to take a look at:
Assessment of Spatial Normalization of PET Ligand Images Using Ligand-Specific Templates.
J.H. Meyer, R.N. Gunn, R. Myers, and P.M. Grasby. NeuroImage 9:545-553 (1999).
I agree entirely with Paul Thompson's advocation of a template that has the average anatomy of the population
under study. This can be motivated in terms of increasing consistency of the estimated warps. For example,
consider registering a pair of images together (say image A and image B). If A is warped to match B, then one
warp will be obtained. If B is warped to match A then another warp will be obtained. If the first warp
is not identical to the inverse of the second warp, then something is wrong somewhere. By enforcing the
constraint that the first warp should be identical to the inverse of the second (and vica verca), then more accuracy
should be introduced. See:
Consistent Linear-Elastic Transformations for Image Matching.
Gary E. Christensen. IPMI'99, LNCS 1613, pp. 224-237, 1999.
A. Kuba et al. (Eds.) Springer-Verlag Berlin Heidelberg
or a number of papers that evaluate rigid registration algorithms based on consistency of the trasformations.
One way of ensuring this consistency is to warp both A and B to an image that is "half way" between the two.
Then matching one image with the other involves taking one warp along with the inverse of the other and
combining them. Similar ideas hold when matching several images together, where instead of a "half way"
image, there is a template comprising of the average, both in shape and intensity, of the images to be
normalised. See:
Image Registration Using a Symmetric Prior in Three Dimensions
John Ashburner, Jesper L.R. Andersson, and Karl J. Friston.
Human Brain Mapping 9:212 225(2000)
There are many issues with inter-group comparisons of functional data. One of them is about the kind of
differences you are interested in - i.e., the pre-processing you do in order to sensitise your statistical analyses
to particular differences.
If one population shows a higher signal intensity than another, then this may be because there is more signal
per unit of grey matter in one population than the other, or it may be because one of the populations has the
same activity per unit of grey matter, but simply has more grey matter. Grey matter maps are needed in
order to disambiguate the two cases. See e.g.:
Positron Emission Tomography Metabolic Data Corrected for Cortical Atrophy Using Magnetic
Resonance Imaging. C. Labbe, J. C. Froment, A. Kennedy, J. Ashburner and L. Cinotti
Alzheimer Disease and Associated Disorders 10:141-170 (1996)
an approach that has been tried in order to approximately do this within SPM is described in:
Cortical grey matter and benzodiazepine receptors in malformations of cortical development.
A voxel-based comparison of structural and functional imaging data. M. P. Richardson,
K. J. Friston, S. M. Sisodiya, M. J. Koepp, J. Ashburner, S. L. Free, D. J. Brooks and J. S. Duncan.
Brain 120:1961-1973 (1997)
Another issue relates to how warping images of different subjects to the same stereotaxic space induces
volumetric differences in the warped images. Assume that a particular experiment causes exactly the same
activation in all subjects, e.g. an intensity increase of one unit in a volume of 1cc. Suppose that in one group,
the region containing the activation is doubled in volume by the spatial normalisation, whereas in the other
group, the volume remains the same. After smoothing, one group will appear to have activations of
approximately twice the magnitude of those of the other. Is this a desirable result?
This relates to the issue of "modulation" in VBM studies that has been mentioned on this list. For a couple
of recent papers that touch on this, see:
Voxel-Based Morphometry Using the RAVENS Maps: Methods and Validation Using Simulated
Longitudinal Atrophy. Christos Davatzikos, Ahmet Genc, Dongrong Xu, and Susan M. Resnick
NeuroImage 14, 1361-1369 (2001)
and:
Why Voxel-Based Morphometry Should Be Used. John Ashburner and Karl J. Friston.
NeuroImage 14, 1238-1243 (2001)
Many of the issues addressed by the latter article are also relevent to functional imaging studies that involve
comparing one population with another. In order to be complete, I guess I'd better mention the paper that
it was written in response to:
Voxel-Based Morphometry Should Not Be Used with Imperfectly Registered Images
Fred L. Bookstein. NeuroImage 14, 1454-1462 (2001)
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
|