I need to compute SNR in specific brain rois derived from parcelization defined in MNI space. Unprocessed images from which I compute noise are not registered neither to subject nor to MNI space. Which is the quickest way to compute SNR over different shells and rois?
thanks,
Rosella
PS: I have tried with the mrregister command to register noise images to MNI space but as I need a per shell analysis I should use the mrregister for each shell and it takes a lot of time !)
Unprocessed images from which I compute noise are not registered neither to subject nor to MNI space.
How do you get unprocessed images to compute noise that are not in subject space?
Which is the quickest way to compute SNR over different shells and rois?
It depends on how you wish to define SNR. DWI data are not multiple acquisitions of essentially the same image, the variance across which you can use to get an estimate of noise: each DWI volume contains unique contrast. You can look at things like SH fit residuals, but that raises the issue of the order of the SH fit. The best estimate of the noise level we can get at the moment is that produced by the dwidenoise -noise option.
If you have estimates of “signal” and “noise”, then generally mrstats with the -mask option is what you’re looking for. Further, given such quantities would be defined in voxels of a subject-space image, it may make more sense to transform the parcels from template space into subject space rather than the other way around.
I have tried with the mrregister command to register noise images to MNI space
You can’t register noise images. Image registration relies on there being some kind of contrast information in both images that can be spatially aligned with one another; a noise level image should ideally contain no biological contrast at all. You could register some other image and then apply the resulting transformation to a noise image.
as I need a per shell analysis I should use the mrregister for each shell
This would only be the case if the data across shells were not already aligned to one another.
1 Like
My unprocessed images are in a different space from the subject image, indeed they are not registered. I finally decided to compute SNR calculation in unprocessed space not to alter noise with geometrical transformation to another space. However, registering parcelization map of subjects to unprocessed space is not perfect and the resulting ROIs are not precise. How should I do?
here you find a link with an example of unprocessed image, parcelization map in subject space and transformed to unprocessed space, and the relative ROIs.
https://onedrive.live.com/?id=B0C0561E34E9D76F!616&cid=B0C0561E34E9D76F