MultiTensor simulator interface

[oesteban]

An interface to simulate diffusion weighted images (DWIs) using the multi-tensor model included in dipy in a ball-and-stick -like fashion:

• Up to 3 fibers are simulated with the multi-tensor model
• Signal is simulated for 3 isotropic compartments (CSF, GM, WM).

Diffusivity of the isotropic compartments and the single-fiber tensor are taken from:

Alexander et al., Detection and modeling of non-Gaussian apparent diffusion coefficient profiles in human brain data, MRM 48(2):331-340, 2002, doi: 10.1002/mrm.10209

[oesteban]

It'd be nice if @Garyfallidis or @arokem could give this PR a quick look before merging :)

[coveralls]

Coverage decreased (-0.19%) to 70.3% when pulling 6b00c67 on oesteban:enh/DipyMultiTensorSimulate into c1a5ea0 on nipy:master.

[coveralls]

Coverage decreased (-0.17%) to 70.32% when pulling 6b00c67 on oesteban:enh/DipyMultiTensorSimulate into c1a5ea0 on nipy:master.

[mwaskom]

It seems strange that the main code here is going to live in nipype and not dipy?

[arokem]

Well - depending on what you mean by 'main', it is. That's the multi_tensor function, the add_noise function and all the functions being used in generate_gradients. The latter could arguably find a home in dipy.core.gradients, or in dipy.sims.

[arokem]

I don't really know how this usually works in nipype, but it would be good to have clear documentation for what goes into args. There's lots of them!

[oesteban]

We need to use args to make the method 'pickleable' by multiprocessing

[oesteban]

@arokem @mwaskom @Garyfallidis I think there are only a couple of features that may be more appropriate in a PR to dipy:

• Processing full images: parallel processing of many pixels and the simple mask computation.
• Simulation of isotropic compartments. This is actually doable already as @Garyfallidis pointed out.

If there's an insterest in dipy of having these two features, I'll be happy to contribute.

[arokem]

Those both sound interesting. I am curious to hear how much of a gain you got from simulating the isotropic component separately. As pointed out by @Garyfallidis, this is already doable, by passing a set of identical evals, but if there's a substantial gain by doing it separately, it might be worth doing. Concerning your second point, we're planning to have a hangout to discuss parallel processing in dipy, and we'd be happy to have your input on that, see link to scheduling on dipy/dipy#631

[oesteban]

@arokem I'll see to participate in the parallel processing chat. However I'm far from an expert. I just know how to use multiprocessing to map a list to a function, and get the results. I don't have any numbers about the performance gain simulating compartments outside multi_tensor. With multi_tensor the b-vectors and the sticks are projected and scaled by the tensor. However, outside, it is just about matrix operations optimized by numpy.

@chrisfilo @satra: what do you think about merging this as is now, and updating this interface when the features under discussion are added into dipy?

[coveralls]

Coverage decreased (-0.18%) to 70.26% when pulling f7c2ff9 on oesteban:enh/DipyMultiTensorSimulate into 328a8fd on nipy:master.

[coveralls]

Coverage decreased (-0.2%) to 70.23% when pulling fc3f27c on oesteban:enh/DipyMultiTensorSimulate into 328a8fd on nipy:master.

[oesteban]

@arokem @Garyfallidis I finally integrated the simulation of isotropic compartments using multi_tensor. However, resulting MD in these regions is lower than that of single-fiber regions. FA doesn't look bad, though. What could I be doing wrongly?

[coveralls]

Coverage decreased (-0.23%) to 70.21% when pulling ef0a9c6 on oesteban:enh/DipyMultiTensorSimulate into 328a8fd on nipy:master.

[satra]

pinging to see where this is at.

[oesteban]

I think that with the last merge with master, everything is ready from my side and all the questions were answered :)