 |
 |
> 1. Even if we can force the volumes match each other, to what > extent should it be done? This is a good question. There is a trade-off between optimising the similarity and keeping the warps as smooth as possible. The amount of trade-off is determined by some hyper-parameters that control the regularisation. With simpler models with fewer parameters, it is often possible to achieve the best balance by doing Bayesian model selection (ie maximising the model evidence for various hyper-parameter settings). Unfortunately, this is currently a bit more difficult for models with millions of parameters. > 1. What are the implications for the final results since this > will be moving so much tissue around? Some regions will expand, whereas others contract. With simple smoothing of warped data, the signal from adjacent expanded and contracted regions may be combined in such a way that those voxels in expanded regions have proportionally more representation in the resulting smoothed data. I've tried to compensate for this in the normalise to MNI space option of Dartel (see manual for more info). The aim was to give pre-processed data that is more similar to what would be obtained from region of interest analyses on the original scans - rather than on the warped scans. > 1. How will this impact inferences regarding for example, > functional activation in the thalamus or caudate? If homologous regions are more closely aligned, the interpretations should be more accurate. Greater overlap of activations in different subjects should lead to greater sensitivity. Not having activations in the wrong place should lead to greater specificity. In Arno Klein's comparisons with human expert labellings, Dartel did reasonably well. Results for some subjects were not as good as for others though, but I suspect that this was due to segmentation failures (Arno did all the registrations and the rest of us were really just advisers and didn't see the actual registration results themselves). Best regards, -John > On Wed, May 5, 2010 at 3:17 PM, John Ashburner > <[log in to unmask]> wrote: > DARTEL > > On Wed, 2010-05-05 at 15:02 -0500, Max Gunther wrote: > > Dear SPMers, > > > > > > Do people have any suggestions for fixing the subcortical > > normalization problems that we have been having? (see > attached jpegs) > > > > > > Notice the misalignment between the midbrains on the zoomed > images at > > the cross hairs. > > > > > > These are done in SPM5 on a Phillips 3T scanner at 1mm > isotropic. > > > > > > So far we have tried: > > > 1. Using segmentation. > > 2. Fixing bounding box issues. > > 3. Manual registration. > > > Any suggestions would be much appreciated. > > > > Max Gunther, PhD > > > > Vanderbilt University Medical Center > > Departments of Psychiatry and Medicine > > Center for Health Services Research > > Nashville, TN www.ICUdelirium.org > > > > > -- > John Ashburner <[log in to unmask]> > > > -- John Ashburner <[log in to unmask]>