Every method has its own advantages and disadvantages, and the ones you
list are only the tip of the iceberg when it comes to devising optimal
registration schemes.
Spatial normalisation within SPM estimates warps based on only about
1000 parameters. Although this is enough represent the overall brain
shape reasonably well, there are a number of relatively small structures
with high contrast edges that require higher frequency deformations to
model them accurately. A prime example of these is the effect of skull
thickness when spatially normalising T1 weighted images. Scalp has very
high intensity in these images. If the registration is based purely
on minimising the sum of squared difference between image and template,
then getting the scalp right will be very important to the algorithm
in terms of minimising the cost function. However, if the subject has
a thick or thin skull, then getting the scalp right will mean that the
brain surface is not right (because the low frequency warps can not
model high frequency differences in skull thickness). This is why the
brain masking step was introduced in order to make the cost function
the weighted sum of squares difference. Voxels outside the brain have
lower (or zero) weight in the registration. The disadvantage of this
is that it is not always as stable as the unweighted registration,
because less information is used. Therefore, the bottom line is that
you should probably use the brain masking if the scalp has high intensity
in your images, but don't use it if the scalp is hardly visible (as in
EPI and most PET images).
Some people choose to skull strip their images (and templates) before
spatially normalising them. This removes the confounding information
from the scalp, but retains the information about the edge of the brain.
This is not the default action within SPM, as skull stripping is either
very tedious to do manually, or difficult to do automatically on images
with a wide range of different contrasts.
A related issue is to do with the ventricles. Again, these are small
structures, with a high contrast between CSF and white matter (on T1
weighted images). Getting the edge of the ventricles correct is
important to the spatial normalisation algorithm. However, for
functional imaging you are only interested in getting the grey matter
as correct as possible. One possible option is therefore to spatially
normalise based on segmented grey matter, matching it to the grey matter
probability map in the apriori subdirectory. We have tried this, and it
gives better grey matter registration for the cerebral cortex. However,
for some reason that I will need to look into more closely, the registration
is not quite as good around the cerebellum. Also, for our data it produced
a small systematic difference in the shape of the spatially normalised
images.
I think that distortions of the functional images is probably one of the
bigger problems. Because the functional images are a different shape
to the structurals, it is not possible to get an extremely accurate
rigid body registration between them using the current methodology in SPM.
If this mapping is not accurate, then estimating spatial normalisation
parameters from a structural image, and applying it to the functionals
will also not be very accurate. For these cases, it is probably better to
estimate the warps from the functional images themselves.
There are also issues about FOV, image quality, same contrast in images
and templates etc that I am sure I have mentioned many times. I realise
that this waffle hasn't directly answered your question, but it may have
increased understanding of what the brain masking option does, and
highlighted a few things to think about if anyone is planning to do
some comparisons.
Although not about fMRI, there is one paper out there (that I know about)
that compares spatial normalisation using structural MRI versus that using
functional images. It is by Meyer et al, in NeuroImage 9:545-663 (1999).
All the best,
-John
| I have a question about how one should normalise EPI fMRI datasets.
| There seems to be (at least) three possible approaches, and I wonder if
| there is a consensus on which is best (or least worst....)
|
| 1) Take a set of high-resolution T1 weighted slices in the same plane
| and with the same FOV as the EPI slices. Normalise this to the T1
| template and then apply the normalisation parameters to each EPI volume.
|
| 2) Take a high-resolution multi-shot EPI image, again with exactly the
| same parameters as the single -shot functional EPIs, normalise this to
| the EPI Template and then apply the normalisation parameters to the
| single-shot functional EPIs
|
| 3) Simply normalise the single-shot functional EPI (usually the mean
| image from the realignment) to the EPI template.
|
| In spm99b, option (3) appears to work really well, so long as the
| spm_defaults are modified so that the BrainOnly option is disabled, and
| it has the advantage of not requiring any other acquistions. It also has
| the advantage that the spatial distortions inherent in single shot GE
| EPI are corrected for. However, there is an assumption here that those
| distortions *should* be corrected. Surely there are some artefacts on
| such images which are dropouts and so that particular region should not
| be pulled out to match the edge of the template.
|
| We have done a quick comparison of (1) and (3) on two of our group
| studies. On one of our studies we obtained more significant activations
| in the final statistical analysis if we used option (3) (i.e.
| normalising the single shot EPIs directly) compared to (1). This was
| initially a surprise as we naively expected only the location of the
| clusters to change. We then came up with the idea that probably option
| (3) was bringing the group data into better register with each other -
| hence the better statistics.
|
| Of course, it had to happen, in the other study we looked at, Option (1)
| gave better final statistical results than Option (3).
|
| Any thoughts? I'd be really interested to know what other groups are
| doing, and whether anybody has done a more rigorous comparison of the
| various methods.
|
| Thanks in advance,
|
| krish
|
|
| --
| Dr Krish Singh, ([log in to unmask])
| Magnetic Resonance and Image Analysis Research Centre, Liverpool
| University
| Pembroke Place, Liverpool, L69 3BX, UK. Tel 0151 7945645. Fax 0151 794
| 5635
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|