MRtrix includes a command
dwidenoise, which implements DWI data
denoising and noise map estimation by exploiting data redundancy in the PCA
domain ([Veraart2016a] and [Veraart2016b]). The method uses the
prior knowledge that the eigenspectrum of random covariance matrices is
described by the universal Marchenko Pastur distribution.
Image denoising must be performed as the first step of the image-processing pipeline. Interpolation or smoothing in other processing steps, such as motion and distortion correction, may alter the noise characteristics and thus violate the assumptions upon which MP-PCA is based.
Typical use will be:
dwidenoise dwi.mif out.mif -noise noise.mif
dwi.mif contains the raw input DWI image,
out.mif is the denoised
DWI output, and
noise.mif is the estimated spatially-varying noise level.
We always recommend eyeballing the residuals, i.e. out - in, as part of the quality control. The lack of anatomy in the residual maps is a marker of accuracy and signal-preservation during denoising. The residuals can be easily obtained with
mrcalc dwi.mif out.mif -subtract res.mif mrview res.mif
The kernel size, default 5x5x5, can be chosen by the user (option:
For maximal SNR gain we suggest to choose N>M for which M is typically the
number of DW images in the data (single or multi-shell), where N is the
number of kernel elements. However, in case of spatially varying noise, it
might be beneficial to select smaller sliding kernels, e.g. N~M, to balance
between precision, accuracy, and resolution of the noise map.
Note that this function does not correct for non-Gaussian noise biases yet.