I understand a bit better now. Your results are probably what would
be expected in that the smallest deformations are likely to be the
most similar to each other. Dartel is less heavily regularised than
the old basis function approaches, so the differences between
deformations estimated using different data are likely to be smaller.
Your approach to validation will therefore give some information about
registration accuracy and reproducibility, but it does not tell the
whole story. For example, consider an approach that involves not
registering them. This would result in identical deformations (ie
identity transforms) for both images, resulting in the least possible
difference between them. It wouldn't mean that it is necessarily a
good approach though.
2011/5/3 Pablo Ripollés Vidal <[log in to unmask]>:
> Thank you very much for your quick answer, Dr Ashburner.
> It is the same subject with and without a simulated lesion. The ground truth
> its the estimated deformation of the subject without the lesion.
> Best regards,
> 2011/5/3 John Ashburner <[log in to unmask]>
>> I don't quite understand what you are doing. Are these T1s from two
>> different subjects? Do you have some form of ground truth with which
>> to compare the estimated deformations?
>> Best regards,
>> On 3 May 2011 18:39, Pablo Ripolles <[log in to unmask]> wrote:
>> > Hello SPMrs,
>> > I am trying to compare deformation fields from 2 different T1s after
>> > being normalized with DARTEL and Unified Segmentation. After unified
>> > segmentation has run I get the deformation field from the .mat with the
>> > deformation utility. I do the same after Create Dartel has finished getting
>> > the def field from the flow field. Then I calculate the root mean square
>> > difference from voxels in the defformation field from image 1 and image 2
>> > first for U.Segmentation and then for Dartel. What I want to see is which of
>> > the 2 algorithms performs better.
>> > From recent results in papers I expect DARTEL to outperform unified
>> > segmentation (Klein et al, 2009) but this only happens when I take the mean
>> > rms of all the voxels in the deformation fields. When I calculate the mean
>> > root square difference for only the voxels from the defformation field which
>> > belong to brain, unified segmentation outperforms DARTEL. For def fields
>> > obtained from unified segmentation I use the brainmask provided with SPM8 to
>> > use only "brain voxels" and for def fields from DARTEL I use a mask obtained
>> > from the Template calculated by the algorithm .
>> > As unified segmentation uses a small deformation approach and DARTEL a
>> > large deformation framework, are the results from de mean square roots
>> > comparable?Am I doing something wrong?
>> > I would really appreciate some help.
>> > Thanks to everyone in advance.
> Pablo Ripollés Vidal
> Ph.D Student
> Cognition and Brain Plasticity Unit http://www.brainvitge.org/
> IDIBELL [Bellvitge Biomedical Research Institute]
> L’Hospitalet de Llobregat, Barcelona, 08097,Spain