I have a questions concerning the use of dwidenoise when Prescan Normalize is turned on. dwidenoise requires that no preprocessing is performed on the data, but I am not sure if prescan norm would be a problem. To provide some background, it removes the brightness changes across the images that are due to the spatial positioning of the receive coils - and it appears to be a basic mathematical step - just a subtraction on a voxel by voxel basis, that doesn’t affect SNR (according to siemens).
I have tested dwidenoise on functional EPI data with and without prescan norm (what I had handy) and the noise maps are slightly different, though no structure is apparent. The resulting timeseries is nearly identical (vast majority of brain >0.95 correlation), thought there are slight changes in a voxel here and there.
Is there a clear theoretical reason that dwidenoise would fail or be incorrect on data that was scaled on a voxel by voxel basis in a way that doesn’t alter voxel SNR?
I think it is fine to use prescan normalisation, provided Siemens calibrates the receive field once and then keeps it static for the duration of the fMRI scan. As I understand it, this is the case and the acquired 4-D fMRI data is simply divided (voxel-wise) by the 3-D receive field.
Technically, dividing by a non-constant image breaks the MP-PCA assumption of a constant noise level. (You are right that it would not affect S/N, but it does affect N.) However, because of the coil profiles the noise level is already not constant throughout the brain, which is why MP-PCA denoising operates in local patches around each voxel where it is approximated as such. Therefore, if the prescan normalisation image is sufficiently smooth (relative to the patch size), it is not violating this underlying assumption more than it already was.
It is reassuring that you have compared both and that the denoised data comes out nearly identical. The noise level maps should be related multiplicatively by the receive field, i.e., the ratio between both noise maps should correspond to the the prescan normalisation factor.
I hope this helps,
It appears that your understanding is correct, based on the image I am examining.
I simply subtracted the prescanned norm data from the original data and the result was a smoothly varying field, brighter in areas closer to the coil. The FWHM (estimated via AFNI’s 3dFWHMx function) is ~63 mm, about a factor of 10x larger compared to the original data, for example. very very smooth.
The difference image was not identical across time, but varied by one here and there. Not sure what would have caused that, perhaps rounding at the level of image reconstruction. My testing data has a very low range, compared to typical 32 Channel data, so that may be to blame.
Thank you for the through answer, if I dig up some other, 'better 'data where I have collected both prescan norm and unaltered data I will update this post.
I did a bit more digging with better data - and examined the the ratio between the noise maps. It is nearly identical to the result of dividing the unaltered data by the prescan normalized data. I think this confirms your latter point.
This second dataset, which is whole brain, 32 channel coil with dramatically different parameters has otherwise identical findings compared to my first set. the correlation between the two denoised time series are higher, in fact (most >0.99). The only difference is in the difference image from (unaltered - psn), which varies as a function of movement, though is also smooth. I believe the ratio is more informative.
So, to wrap up, denoised data looks (even more) similar, the ratio of noise maps looks like the prescan normalization map and it is all smooth. If I get really motivated, I’ll test again with diffusion data, instead of BOLD.