Thanks Helmut. Quite helpful. My admittedly limited previous experience with deformation maps (which is what I understand the y_anat.nii files to be) lead me to expect... something else.
I'll try the manual registration correction. Maybe I can tweak those people a bit. It is a big data set (177 subjects for a task study) so it has been a fair amount of work : )
I generally have full coverage in the functional data, but sometimes the edge of the brain is on the top slice of the fMRI image is the top edge of the brain or scalp. The registration does not seem to deal well with these cases, even though we still have the whole brain.
Colin
-----Original Message-----
From: MRI More [mailto:[log in to unmask]]
Sent: September-29-16 12:04 PM
To: [log in to unmask]; Colin Hawco
Subject: Re: normalization issues
Dear Colin,
The "gradient" pattern is what the y_ usually look like. Maybe you accidentally looked at y_ files from some subjects and u_ files from others?
> Also the offset images are more likely to have a small misregistration with the anatomical T1, but all attempts to get coregister to do a better job have not produced improvements.
You could still correct these images manually. It's subjective, but well, in that case it makes more sense than relying on objectively but poorly registered images.
> The deformation fields (the y_anat.nii file) do not seem correct to me.
Not sure whether I understand this correctly. Is it the anatomical images which are "incomplete"? In case some parts of the brain are cut off and/or the brain extends close to the boundary (thus, parts of the skull and dura missing) there's probably not much to do except if you have another image that covers the whole head, which you could coregister the "incomplete" data onto and then use the "complete" ones for segmentation purpose. Of course, you could still try to work with tissue priors/templates that are cut off at the top to the same extent, but this would be time-consuming and probably quite subjective. However, if the c1, c2 are alright (not incomplete) I would suggest to Dartel register these together (probably combined with an additional transformation into MNI space). As it's just the c1, c2 (well, the rc1, rc2) it wouldn't matter whether the other tisse class images are "incomplete" or not. This should lead to a undistorted normalisation, but as stated, it requires unaffected c1, c2, which might not be the case.
Or is it the functional images which are "incomplete"? In that case the problem should be solved by manually correcting the coregistration onto the anatomy. You might still encounter something like in the attached figure, where parts of the brain are extended into regions where no data was available, but this can be solved by applying an explicit mask to the first-level models.
Hope this helps a little
Helmut
|