NF - Particle Filtering Tractography

[gabknight]

This PR add a new tissue classifier (CMC) and a new tractography algorithm (PFT). Both methods are described in [Girard et al., NImg, 2014].

• CMC - new tissue Classifier. CmcTissueClassifier used probabilisties and the tissue partial volume fraction maps to determine when the tracking stops. As it is very similar to ActTissueClassifier, both classes now extend an a new abstract class ConstrainedTissueClassifier.

• PFT - new tractography method. ParticleFilteringTracking is inherited from the LocalTracking class. The particle filter is used to find a valid streamline path when the streamline being traced reach the state INVALIDPOINT.

  • An example is giving in doc/examples/particle_filtering_fiber_tracking.py

@nilgoyette found potential optimisation for the cumsum(...) function, called several times in dipy.tracking.local.localtrack._pft(...) and dipy.direction.probabilistic_direction_getter.ProbabilisticDirectionGetter.get_direction(...). A PR should follow :).

@MrBago I would be very glad to get your feedback on the way I actually structured the implementation of PFT in dipy. PFT is the first tracking algorithm in dipy that is not a different direction getter as such. PFT uses a probabilistic direction getter, and needs to know the tracking state from the tissueClassifier at every steps and use it change the reconstruction of the streamline. I thus choose to extend the LocalTracking class with the ParticleFilteringTracking class and added the _tracker(...) function to the LocalTracking class. ParticleFilteringTracking returns the pft_tracker cython function, LocalTracking returns the local_tracker cyhton function. Both use the same _generate_streamlines(...) function from LocalTracking.

@Garyfallidis, @arokem, @skoudoro, @jchoude, I would be very much interested to hear your thoughts on this implementation. On the CC exemple, PFT is ~4.5x slower than ProbabilisticDirectionGetter LocalTracking, but outputs 3x more streamlines connecting the GM (i.e. not stopping in the WM or CSF). PFT has a seed to streamlines ratio of ~93% while the same algorithm With localTracking has a ratio of ~29% (for CC seeding of the example).

References

Girard, G., Whittingstall, K., Deriche, R., & Descoteaux, M. Towards quantitative connectivity analysis: reducing tractography biases. NeuroImage, 98, 266-278, 2014.

[skoudoro]

Since #1342 is merged, can you rebase this one @gabknight? Thanks !

[nilgoyette]

@gabknight I don't remember what we said... you rebase and I open another PR for some enhancements, or I rebase and enhance in this PR?

[gabknight]

@nilgoyette @skoudoro I will rebase this.

@nilgoyette I will also update the particle filtering part following #1342. It would be ideal if you can then have a look to the whole PR for further optimisation.

[gabknight]

I screw thing up here.... I open a new PR #1384 with the good commits