Preferred DWI Denoising pipeline

Dear All,

I am working on DWI images and want to understand whether the selection of the files for denoising will impact the quality of the final results. I have data of 80 volumes of DWIs (25 volumes of b1000, 24 volumes of b2000, 24 volumes of b3000, and 7 volumes of b0) in PA direction. In addition, I have 2 volumes of b0 in AP and 2 volumes of b0 in PA direction of the same subject in the same run.

Now, I want to understand what should be the ideal pipeline of preprocessing. Following are my approaches:

  1. Denoise AP b0 (2 volumes), PA b0 (2 volumes), and DWI (80 volumes) separately.

  2. Denoise AP- PA b0 pair (4 volumes) after concatenation and separately denoise DWI (80 volumes).

  3. Denoise all the available volumes i.e. AP- PA-DWI (84 volumes) altogether.

  4. Denoise AP b0 (2 volumes) and PA-DWI (82 volumes) separately.

  5. Denoise only DWI (80 volumes), and process AP-PA b0 pairs (4 volumes) directly into fsl topup without performing denoising.

If anybody has tried any of these approaches and achieved significant results, then let me know.

@Mrtrix_developers if you have come across any such query earlier in the past then connect me to that discussion. Also, give me your valuable inputs on each of the approaches, what can be the possible disadvantage and advantage of using each of the approaches.

Thank you in advance for giving your time.

Muskan Khetan

Hi Muskan,

Excellent question, thanks for formulating it so clearly. I have experimented with several of these options and I find that approach 5 works best, at least for a “standard” protocol with a AP-PA pair as you’re describing. You can pass these AP-PA b0 images separately to dwifslpreproc using the -se_epi option.

Some thoughts about why we prefer option 5:

  • Options 1 and 2 won’t do much of denoising on the AP/PA b0 volumes because the no. volumes (4) is too small.
  • Options 3 and 4 can potentially be problematic because the AP-PA geometry shift increases the local rank of the signal and may as a consequence lead to an underestimation of the noise level.
  • Option 5 distinguishes between the field map estimation, based on the AP-PA images, and the denoising & geometry correction of the dMRI data that will be used in further analyses. The field map estimation is strongly regularized so not too sensitive to noise; besides, the b0 images are of high SNR anyways. Its output is then used as an input to the main data processing pipeline, in which in this case the noise level estimation is done entirely independently.


Hi Dr. Daan,

Thank you so much for giving an explanatory reply on the use of each approach. This will really help me in my work.