Dear Philip,
How far outside the brain and into the white matter do the clusters extend?
If it is just a small amount then you can ignore it, as the registrations will not be perfect and so there is often a small amount of inaccuracy like this.
If it is a large amount then that would point to there being a bigger problem with the registrations, and I would look carefully at how the anatomical images are aligned (not just at the overlays of the activations). It could either be the functional to structural or the structural to standard registrations that need attention, so make sure you check both. The functional to standard registration is simply made up of the other two and is not independent, so just check the other two.
All the best,
Mark
On 23 Jan 2014, at 12:57, SUBSCRIBE FSL Philip Lindner <[log in to unmask]> wrote:
> Dear FSL experts,
>
> I am trying to investigate group differences in resting-state networks using Melodic and DualRegression (DR), but have encountered some difficulties that I have been unable to solve. In a nutshell, performing DR after Melodic turns out p < 0.05 (FWE-corrected) significant clusters located in white matter regions and outside the brain (along with some cluster inside hypothesised grey matter regions). I have understood that getting significant cluster outside of the component map in question is not a problem, but finding these clusters in WM and outside of the brain worries me. I have re-run the analyses several times, changing some variables in the Melodic-stage, but always with the same kind of problems after DR.
>
> My first guess was that registration to T1 (1mm, using BBR) and standard space (MNI-1mm or 2mm, with 12 DOF) had gone wrong. If I superimpose the melodic_IC.nii.gz file on the standard image, there are indeed some parts of the component maps that stretch into WM and outside the brain (respectively) – depending on how I threshold the melodic_IC.nii.gz file, of course. I have however used the transformation matrices from the registration phase to transform seed masks from standard and high-res space to fMRI-native space, and these transformations turn out OK. In Melodic, not including registration to T1 and changing the reference standard image does not help. I use the standard settings but usually extract 20 components and re-sample to 2mm. N = 30+22. In DR, there is not much to change in the settings themselves.
>
> I cannot continue with my study until I have resolved this issue and have reliable results. What do you think could be the error here? Has anyone come across the same error and managed to solve it? I have found one way of alleviating the problem and that is to threshold the melodic_IC.nii.gz file (at some arbitrary level) before running DR on it. What do you think about this option?
>
> Thank you very much for your help!
>
> Regards,
> Philip
> PhD student, Karolinska Institutet
|