Reconstruction of bundles of interest

When a researcher is interested specifically in one or a small number of pathways of interest, there are multiple strategies that can be employed to reconstruct and interrogate such. Each has their pros and cons.

“Targeted tracking”

  • Conventional approach

  • Constraints based on the pathway of interest are applied during streamline generation, ie. tckgen

  • Typically define a seed region from which to commence streamlines, a target inclusion region that they must reach, and potentially additional inclusion / exclusion regions to constrain where the streamlines can & cannot go on their way there

  • Common to use -seed_unidirectional to track only in one direction from the seed point

  • Can optionally use -stop to cease streamline propagation the instant it touches the final inclusion region

  • Common to perform twice, once seeding from A and propagating to B and once seeding from B and propagating to A

    • tckedit can be used to merge multiple track files


  • Can produce a dense reconstruction even for minor pathways, as streamlines are not seeded in locations far away from the bundle of interest


  • Can be difficult to diagnose cases where you don’t obtain expected results in either number or trajectory of streamlines

    • Using -info provides summary statistics that can help
    • In extreme cases, uncommenting this line and re-compiling will yield images that provide additional insight into where streamlines went and why they were rejected
  • Potential trap if using a combination of:

    • ACT, which terminates streamlines precisely at the isocontour between WM and cortical GM
    • ROIs that only include voxels that are primarily cortical GM

    , as streamlines may be terminated just prior to intersecting the ROI and therefore be omitted from the output

    • For now, dilating such ROIs to include the GM-WM interface may help; better technical solutions are in the pipeline.
  • Have to re-generate streamlines from scratch if criteria change

  • Can’t apply calculations that necessitate use of a whole-brain tractogram (e.g. afdconnectivity -wbft, tcksift, tcksift2; see explanation)

“Tract selection”

  • Similar to targeted tracking, but breaks the reconstruction into two steps rather than all-in-one:
    • Perform whole-brain fibre-tracking (tckgen)
    • Apply constraints that isolate from the whole-brain tractogram only those streamlines corresponding to the pathway of interest (tckedit)


  • Can modify pathway selection criteria and re-apply without re-generating streamlines

  • Can modulate application of criteria (e.g. add / remove ROIs, erode / dilate ROIs) to see which streamlines are “at the edge” of selection

  • Can apply calculations that necessitate use of a whole-brain tractogram


  • Still potential trap RE: combination of ACT and GM-based ROIs from “targeted tracking”

  • Won’t get as many streamlines in the selected pathway as what you would obtain using “targeted tracking”, as more computation time is spent generating streamlines that aren’t in the pathway

“Connectome subset”

  • Premise is that the pathway of interest is a specific edge in the connectome, based on two endpoint regions that are members of a parcellation

    • It’s still possible to use this approach even in the absence of a whole-brain parcellation; but it involves integrating the regions of interest into a kind of “dummy” region; will add details of how to do this if there is adequate demand)
  • Sequence of steps:

    • Generate all data requisite for connectome construction, including whole-brain tractogram
    • Generate connectome with tck2connectome, but additionally use -out_assignments option
    • Use connectome2tck to extract streamlines corresponding to edge of interest


  • Can apply calculations that necessitate use of a whole-brain tractogram

  • Includes heuristic for assigning streamlines to parcels that is not strictly dependent on an intersection between streamline and parcel; is therefore more robust when used in conjunction with ACT

  • Potential applications of the additional data generated, of which the pathway of interest was just a subset


  • Heuristic for assigning streamlines to parcels is imperfect, can result in erroneous assignments in some circumstances

  • Doesn’t directly permit more complex criteria on streamline trajectories

    • Though these can be applied with tckedit subsequent to the connectome2tck call if desired
  • Won’t get as many streamlines in the selected pathway as what you would obtain using “targeted tracking”, as more computation time is spent generating streamlines that aren’t in the pathway

1 Like

Hi everyone,

I have two questions about “Reconstruction of bundles between ROIs pair”

  1. If we manually draw two ROIs in MRtrix and perform fiber tracking, however, it doesn’t exist any white matter connections between this ROI pair in neuroanatomical knowledge. Are the streamlines between this ROI pair false positive?
  2. And is the number of streamlines between this ROI pair (no anatomical connections) very low compared to the ROI pair with truly anatomical connections?

Thank you for your time!


Hi Robert,
Thank you for your patient answer!

Could you tell me where to find the “better technical solutions are in the pipeline”?


Hi Rob & team,

Thank you for this great summary & all the helpful information available on this forum.

I have a question related to this specific “con” related to the Tract Selection approach:

I have recently run both the targeted tracking and the tract selection approaches to generate (for the first approach) / select (for the second approach) streamlines between two ROIs, both of which are 6mm spheres (non-linearly transformed to subject space) (the small ROIs are selected based on a previous project, however shifted slightly to cover more of the gmwm interface).

I have indeed found exactly what you suggest here, that there are substantially fewer streamlines by using the tract selection approach.

Apologies if it’s one of those questions that doesn’t really have an answer / or if it has already been answered elsewhere on the forum, however I am wondering if there is a rule of thumb regarding the minimum number of streamlines between two ROIs where you would calculate a sum of streamline weights that would be considered “reliable”? For example, if across my subjects there is a mean of 20 streamlines selected between the two ROIs, with some subjects only having 4 or 5 streamlines in this connection, would this number be considered too low for a reliable calculation?

Thanks in advance for any advice,