I have a question.
Is there any parameter that can be set for the MPPCA algorithm in the command “dwidenoise”?
Like we can change the Gaussian smoothing parameter or the search window radius of NonLocal-Means denoising algorithm to get different denoising degrees of NonLocal-Means algorithm?
The only parameter you can change in that algorithm is the size of the patch, using the -extent option. This can make quite a big difference in some circumstances, but it isn’t really about controlling the amount of denoising – it’s about providing enough data to the algorithm to figure out what the noise level is, if that makes sense.
Thanks. Although I have the following picture in your literature “Diffusion MRI Noise Mapping Using Random Matrix Theory”, this confirms your answer to me : The only parameter you can change in that algorithm is the size of the patch. I also have a question, according to your description, is there some parameters that I cannot change? If there are these parameters that we can’t change, what are their default values?
I’m bothering you because I really want authoritative confirmation. Thank you very much for your patient answer.
Hi, I have two questions about dwidenoise command:
1、Can I denoise Rician noise using dwidenoise command(MP-PCA algorithm) directly? Or I must do VST(variance-stabilizing transformation) on DW images with Rician noise before I use MP-PCA?
2、Are there any articles that describe this in detail? Why? “By default, the command will select the smallest isotropic patch size that exceeds the number of DW images in the input data, e.g., 5x5x5 for data with <= 125 DWI volumes, 7x7x7 for data with <= 343 DWI volumes, etc.”
You can denoise the data directly, if you assume Gaussian noise. If you wish to correct Rician bias, it is indeed advised to first apply Rician noise correction and then denoise the output. Mind that Rician correction also need an estimate of the noise level, so you could first run dwidenoise to get the noise level, then apply Rician correction in the dMRI data (not denoised), and then apply dwidenoise again on the Rician-corrected data.
The main reference is Veraart et al. (2016). The choice of the patch size is heuristically defined. In our experience, this is a good default for most data, but feel free to experiment with this for your data.