NF: DKI+ constraints

[tomdelahaije]

This request implements the DKI+ constraints proposed in https://doi.org/10.1016/j.neuroimage.2019.116405. The constraints can be used by setting fit_method='CLS' (constrained ordinary least squares) or fit_method='CWLS' (constrained weighted least squares). The implementation can be tested by following the DKI example in the documentation (https://dipy.org/documentation/1.4.1./examples_built/reconst_dki), and running for example

dkimodel_plus = dki.DiffusionKurtosisModel(gtab, fit_method='CWLS')
dkifit_plus = dkimodel_plus.fit(data_smooth, mask=mask)

It is recommended to use the MOSEK solver for speed by setting cvxpy_solver='MOSEK', cf. https://docs.mosek.com/latest/install/installation.html for installation.

The implementation of these constraints is essentially identical to the implementation provided by #2531 for MAP-MRI.

One point: the SCS solver that CVXPY uses by default is more stable and gives more accurate results when the b-values are given in natural units [ms / micron^2], essentially dividing the b-values by 1000 when they are given in standard units of [s / mm^2]. Is there a recommended way of using these natural units when setting up a gradient table, or should I just try to rescale the gradients in the fitting function?

[pep8speaks]

Hello @tomdelahaije, Thank you for updating !

Cheers ! There are no PEP8 issues in this Pull Request. :beers:

Comment last updated at 2022-09-09 13:06:11 UTC

[codecov[bot]]

Codecov Report

Merging #2609 (ca102ce) into master (f9dd0f5) will decrease coverage by 0.41%. The diff coverage is 80.72%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2609      +/-   ##
==========================================
- Coverage   83.92%   83.51%   -0.41%     
==========================================
  Files         132      135       +3     
  Lines       18479    18994     +515     
  Branches     3047     3106      +59     
==========================================
+ Hits        15508    15863     +355     
- Misses       2228     2359     +131     
- Partials      743      772      +29     

Impacted Files	Coverage Δ
dipy/core/optimize.py	79.06% <0.00%> (+2.59%)	:arrow_up:
dipy/reconst/utils.py	96.00% <ø> (ø)
dipy/viz/plotting.py	7.22% <0.00%> (-0.18%)	:arrow_down:
dipy/reconst/dki.py	94.49% <85.71%> (-1.90%)	:arrow_down:
dipy/reconst/dki_micro.py	91.60% <100.00%> (-0.70%)	:arrow_down:

... and 9 files with indirect coverage changes

[skoudoro]

Hi @tomdelahaije,

Can you give us an update on this? thank you!

[tomdelahaije]

Hi @tomdelahaije,

Can you give us an update on this? thank you!

Hi @skoudoro

Sorry for the delay, there were some developments on the theory side of this that required some work. I'll get back to you with an update later today.

[skoudoro]

Ok, no problem, thank you for the feedback.

[kerkelae]

@tomdelahaije, in our lab, we would like to use your method to calculate kurtosis maps. Is this code reliable or would you recommend using another implementation? Thanks for the great work!

[Garyfallidis]

@tomdelahaije can you update this PR? Many people want to use it asap. We can also highlight it during the upcoming workshop http://dipy.org/workshops/latest

[skoudoro]

PR updated, thanks @tomdelahaije for your help!

It would be great if @kerkelae or @RafaelNH or @arokem could look at this and give feedback.

I think I might go ahead and merge this nice features tomorrow if there are no additional comments.

[Garyfallidis]

Congratulations @tomdelahaije !!! Another great + technique merged!

[Garyfallidis]

@kerkelae and all please start using this method and provide feedback to @tomdelahaije !

[kerkelae]

@Garyfallidis I've been using the code from @tomdelahaije's own branch for a while now and everything has been working well! Thanks for making it a part of DIPY. This is a really valuable contribution.