| I am analyzing a large SPECT perfusion (99mTc-ECD) normal dataset
| (20-80 yrs), looking at possible age and gender effects. All
| individuals also received MPRage T1 imaging. Images were normalized
| using the individual's MR images with non-linear parameters (7x8x7
| basis functions = default), to a voxel size of 3x3x3 mm3.
|
| My questions are as follows :
|
| 1. Is this non-linear default the optimal selection or does an
| (empirical?) optimisation exist for the number of basis functions
| (i.e. similar to Stamatakis et al, Nonlinear spatial normalisation
| of SPECT images with SPM99, HBM2000 poster 527). In their abstract
| at HBM2000, Cardenas et al. concluded that especially for subcortical
| regions SPM warping was good but suboptimal. Would one expect that a
| higher number of basis function provide improved central nuclei
| registration ?
Increasing the number of basis functions can improve the registration,
but it can sometimes make things worse, depending how good the
registration model is for your data. The registration model is not so
good if the brain images contain large lesions, or where the contrast
in the images differs from that of the template. In cases like this,
using extra parameters may result in the normalisation procedure
over-fitting the data. To prevent over-fitting, you can either
reduce the number of parameters, or increase the amount of
regularisation. I prefer the latter option.
If you have good data, then you should find that increasing the number
of parameters improves the quality of the spatial normalisation.
More parameters allow a wider variety of different brain shapes to be
modelled.
|
| 2. By linear regression analysis versus age I found several regions
| with decreased perfusion as a function of age. In order to estimate
| the influence of structural changes and atrophy, a voxel-based
| morphometry was done on the segmented grey matter images with 12 mm
| smooth.
| This was done on the (non-linear) normalized T1 data. Approximately
| the same regions were found as in the perfusion analysis.
|
| 2a. Is this a correct approach or should affine-normalized data be
| used for segmentation ? Can the relative number of grey matter voxels
| (Nvox) from the segmentation output be used as parameter to estimate
| global atrophy changes with age or would this be biased by the
| non-linear normalisation ?
I think this is the correct approach, but I know that some groups just
use affine normalisation for their VBM. It may also be worth considering
what the effects of spatial normalisation are on the data. Some regions
grow, whereas others shrink during this procedure. This can influence the
results of VBM, and other analyses that involve comparing different subjects.
The total number of grey matter voxels can be counted from an image
constructed from the normalised segmented image multiplied by an image
of Jacobian determinants derived from the deformations estimated during
the spatial normalisation stage. Jacobian determinants can be obtained
by the routine attached to my email of 13th July.
| 2b. Should the (MR) voxel size be comparable to the original dataset or
| is 3x3x3 mm3 OK since before analysis 12 mm extra smoothing is performed ?
If you are segmenting your spatially normalised images, then it is better
to use higher resolution data to reduce the effects of partial volume.
Otherwise, it should be relatively OK to use lower resolution spatially
normalised data. One possible disadvantage may arise if the resolution
of the data is reduced too much though, and that relates to the frequency
of sampling of the original data. If the original data contain very high
spatial frequencies, then this information is lost when you sample at
a lower frequency, which can introduce aliasing effects (you lose some
signal if you only sample every few voxels). In such cases, it can be
better to smooth the data slightly before resampling.
| 2c. Can I conclude from a comparison of the regression parameters that,
| in case of non-significantly different coefficients, some of the
| perfusion decreases are artefacts since they are structure-based ?
I don't think you can say that there is definately a reduction in perfusion
per unit of grey matter, as the results you see could be attributed to
a reduction in grey matter with the same amount of perfusion.
| 2d. Is inhomogeneity correction necessary when segmenting MR T1 images
| of different age groups since age-related signal contrast differences
| exist (Davatzikos et al. "Effects of aging on the MR signal
| characteristics of brain tissue" (poster 660)) or is this irrelevant ?
The inhomogeneity correction models smoothly varying intensity changes
rather than intensity differences between tissue types. However, one
effect of aging and the associated decrease in grey/white contrast
may be systematic bisases in the ability of segmentation methods
to distinguish different tissue types. Some of the results seen
by VBM analyses could be due to different contrasts rather than different
amounts of grey or white matter. Within the Bayesian scheme used
by the segmentation in SPM99, classification of regions of poor contrast
will be based more heavily on the spatial priors used by the segmentation.
|
| 3. A more rigid approach would be a complete regional partial volume
| correction, but I am not aware of (compatible) software to investigate
| this ?
As far as I know, there is no really proper way of doing this with SPM.
The closest possible approach would be to look for interractions between
grey matter concentration and perfusion, as described in:
M. P. Richardson, K. J. Friston, S. M. Sisodiya, M. J. Koepp, J.
Ashburner, S. L. Free, D. J. Brooks and J. S. Duncan (1997).
"Cortical grey matter and benzodiazepine receptors in malformations
of cortical development. A voxel-based comparison of structural and
functional imaging data". Brain 120:1961-1973
The approach does make a number of assumptions, which need to hold
in order to acheive correct results. A more rigorous analysis
would probably require the perfusion images to be divided by
corresponding images of grey matter concentration. However, these
ratio images are extremely badly behaved statistically, so some
kind of an iterative reweighted least squares method would probably
be needed.
Best regards,
-John
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|