Clustering streamlines?

Is there a way (software, algorithm, or script) to parcellate cortical areas based on their connectivity similarity. For example, group neighboring voxels in the gm/wm interface if their streamlines have similar trajectory and endings. I remember vaguely there was some work in this direction some years ago. Does anyone knows if this idea has been concretized with easy to use tools on MRtrix outputs?

Thank you for any help.

Hi Dorian,

I spent half of my PhD trying to do something along these lines; but I never published the methods beyond ISMRM abstracts, nor added them to MRtrix3, as I was never convinced that they would solve more problems than they would raise (it’s a long story; I can send you a copy of my thesis if you want).

I would probably start by clarifying how I tend to treat the nomenclature around such methods, since it has an effect on the subsequent discussion:

  1. Clustering” generally refers to grouping streamlines based on traversing a similar white matter pathway, and so is based on comparing streamlines trajectories in some geometric fashion. QuickBundles is probably the most prominent example; Greg Parker at Cardiff is the other person most recently pursuing this who comes to mind.

  2. Connectivity-based parcellation” refers to grouping voxels / vertices based on having similar connectivity to the rest of the brain, probably most popularised in Tim Behrens’s thalamic parcellation work. So the distinction is: using the parcels in which streamlines terminate at the opposing end to drive the grouping, rather than the streamline trajectories as in 1… There’s still work going on here both in diffusion and fMRI from various places; I won’t try to give a list since I’ll invariably offend somebody, but if you search for “connectivity-based parcellation” rather than “clustering” you might get closer to what you’re looking for.

I have the beginnings of a script lying around somewhere for doing the latter (just within a region of interest; not from whole-brain to whole-brain, which is a much harder problem) that might make it into MRtrix3 one day. But as it stands currently, no there’s no one-button solution for performing either type of processing. Both are still very tough problems to devise robust solutions for, so authors are maybe less willing to release them (or even publish them in my case :sweat_smile:).

Sorry to disappoint :pensive:

Thank you Rob, not disappointed at all, this info is what I was looking for. I would be interested to give a try to your script if you want to share.

My intent, just to be clear, is not to do a research study. We have a practical problem of guiding the therapy on individual patients. We find that some patients benefit when certain areas of the brain are stimulated, all in the same lobe (sorry for not being open, project managers might not be happy about it). The area of stimulation though is a hit and miss. We think DTI can help in this regard, first by providing a connectivity-based parcellation of that lobe (or entire brain if necessary), then by checking if successful applications fall in an area with a similar connectivity pattern. So, my interest falls in the second option.