samuelstjean
commented
5 months ago Hi there Anar,
As for the first point, the noise maps are acquired directly at the scanner, that is you need to plan your acquisition to acquire one of those, it is not an output of anything unlike the sigma/noise standard deviation maps that all the method produces as an output from the data. As such, if you acquire noise maps without an RF pulse (if memory serves me correctly, but @delucaal might remember more than me from the acquisition side), you will indeed only end up mesuring stuff from the background without the object. This slightly underestimates the distribution in a way because noise in pMRI methods depends on what you image also, but that mistake is usually less than mixing the object signal with the noise distribution, since you have a simpler case and thus make less errors.
On the second point, you are correct, we only measure variations from the background, because the available sample data is far larger than on the image point. It's a tradeoff, since if you measure locally, you end up with one distribution at each voxels, and the coil combine will mash that up even more. So you have to apply a small estimator in a window, assuming every voxel in that windows comes from the same distribution. But in your everyday DWI, each bvalue, and each direction, has a different intensity, and thus is likely to arise from a different distribution. This is even worse when you mix in a windows different tissue types, leading to errors in your distribution/estimation. Depending on your application, this might be good enough, or not, it's hard to say without knowing the end goal.
This is where noise maps come in, at least you know you underestimate a bit, but that estimation is far more stable since the underlying distribution is always centered on 0, and not on the tissue intensity value in that one voxel. I actually had ideas to make it more general and working over tissues, but it still needs more pen andp aper work to have something stable and double check my formulas I think, maybe someday!
Hi autodmri team,
Firstly, I'd like to extend my thanks for releasing the code and for your groundbreaking paper. Your work, which frames the domain of diffusion MRI in the context of gamma distribution and other advanced statistical methods, is both insightful and pioneering.
I have been thoroughly reading your paper and exploring the code. While I find the methodology section clear and well-explained, I am encountering some difficulties in fully grasping the results section, particularly in the interpretation of the specific outputs from MPPCA, LANE, PIESNO, and your method.
Could you please provide some additional clarification on the following points?
.nii.gzfile? I have seen in the directory of your datasets, that there is a noise map from SENSE but this is giving noise distributions for the backgrounds, no?autodmripredicts the noise values for the background-only voxels, is that true? I am just wondering how practical or useful can this be if you want to intrinsically understand how noise affects the DWI data. Usually, for quality enhancement MRI protocols, you want to analyze different noise maps and see whether they capture the noise depth properly.Your insights on these matters would be incredibly helpful for my understanding and further exploration of your work. Thank you once again for your contributions to the field, and I look forward to your response.