I was wondering if anyone has suggestions on how to obtain a numerical measure of endpoint distribution that could be used in downstream statistical analyses. For context, we are performing ROI–ROI tractography rather than constructing a whole-brain connectome. The tracts we are studying are relatively small, and in our experience whole-brain approaches are both computationally demanding and somewhat unreliable for reconstructing these specific bundles.
One idea we are considering is quantifying the spatial distribution of streamline endpoints within a target ROI. For example, if we look at streamline terminations within the hippocampus, we would like to derive a metric that captures the proportion of endpoints located anterior versus posterior along its long axis. The goal would be to reduce the endpoint distribution to a continuous measure that could then be related to behavioral variables.
Conceptually, this is motivated by preclinical literature suggesting functional gradients along the hippocampal axis. However, I am aware that diffusion MRI operates at a much coarser resolution than rodent tracer studies, so it is not obvious how well such gradients can be captured in human tractography data. My main questions are therefore: first, whether quantifying endpoint proportions within an ROI in this way is likely to be biologically meaningful, and second, whether there are established or defensible ways to implement such a measure within an MRtrix-based workflow or even using other software you might be aware of.
For reference, we are working with two datasets: the HCP 7T diffusion data (1.05 mm isotropic voxels) and an in-house dataset acquired at 2 mm isotropic resolution. Any suggestions on approaches or best practices for extracting these kinds of endpoint distribution metrics would be greatly appreciated.
I don’t know if I’m the best person to speak about this, but I’m curious what your thought process for looking at the spatial distribution of streamline endpoints is. Tractography is inherently “fake,” since it’s directly dependent on where/how you seed, especially for ROI–ROI (i.e., in whole brain tractography, tcksift2 is used to weight the streamlines based on the FODs, but—to my knowledge—this is not possible for ROI–ROI tractography).
Whether or not this would even be biological meaningful is a good question…as you said, DTI has a very coarse resolution, so even if you had endpoint density metrics, I don’t think it would be very trustworthy.
One idea I have is to potentially rely a little less on tractography for your metrics. Going off your hippocampus example, perhaps you could reconstruct two tracts: one to the anterior part and the other to the posterior. You can then use tckmap to identify voxels traversed by these streamlines in each subject, warp the resulting tract-based ROIs to MNI, average them, warp them back to each subject, and extract FA, MD, etc. This is similar to what I have done in my work.
I hope this is helpful! Again, I do not have much experience on this, so please take my advice with a grain of salt.
In our case, we generated a whole-brain tractogram as well as a regional tractrogram, applied SIFT2 to the combined set of streamlines and then extracted streamlines (with corresponding SIFT2 weights) that had an endpoint in the hippocampus. As far as I understand, whole-brain tractography is important: SIFT2 relies on the relationship between the tractogram and the underlying FODs, so it requires a whole-brain tckgen rather than ROI–ROI tractography to produce meaningful weights.
While doesn’t fully resolve the biological validity question, it does provide a more defensible framework than using raw streamline counts.
Hi Arkiev, we read your paper a bit ago, and it was what inspired me to approach this direction of research.
While SIFT2 and a whole-brain approach seems to be the consensus approach to this idea, we are tracking very tiny white matter bundles, and a whole-brain tractogram doesn’t give us robust enough tracts for us to meaningfully compare endpoint terminations. We’ve reached the determination that careful ROI-ROI tracking is possibly the best way to go about delineating the tracts of interest.
I’d like to know your and the communities’ thoughts on the following ideas we had:
Using LiFE lesion analysis to do something similar to what SIFT2 does to validate streamlines based on the diffusion signal (though given how RAM intensive this is, I don’t think it is computationally feasible to run this for ~170 people)
Get endpoint maps, but then normalize by streamline counts and ROI size so we can quantitatively compare.
If running SIFT2 or a similar model such as COMMIT is absolutely essential to be able to make these endpoint maps something we can use as a measure, could we take our ROI-ROI tract and concatenate it with a whole-brain tractogram and then run SIFT2 on this concatenated whole-brain tractogram and then isolate the streamlines we want?
Thank you for your suggestions, looking forward to hear what you and others think about these potential ideas!
That’s great! If ROI-ROI tracking is providing sensible reconstructions, the methods we used should in theory be replicable. This is in contrast to ROI-ROI tracking followed by SIFT2 alone — my understanding is that, in the latter case, any weights produced would not be meaningful.
As for the ideas:
I’m not sure I fully grasp what is meant by validating streamlines based on the diffusion signal… As for computational requirements, 170 participants isn’t necessarily the limiting factor if jobs can be submitted via SLURM or HPC. The number of streamlines that SIFT2 is applied to is more likely to be the bottleneck, as this will drive RAM and processing time requirements.
Extracted streamlines and their corresponding SIFT2 weights can be used to generate endpoint maps and to make quantitative comparisons — without needing to normalise by streamline count. That said, normalisation by region surface area may be worth considering.
Yes, precisely — at least in theory, and at least for SIFT2 (I’m not sure about COMMIT). The approach worked well with HCP data, though I could imagine running into issues with clinical datasets.
Glad to hear the paper has been helpful! If the work aligns with what your group is doing and there’s interest in collaborating, feel free to send me a DM
Hi Arkiev, thank you for your reply! As for a collaboration, I can bring it up to my PI, I think we’re wrapping up our main projects and the people who do DWI (including myself) are all leaving or will be leaving soon so I am not too sure if that would work out unfortunately
For 1) when I said validate by diffusion signal, I meant using LiFE ( Evaluation and statistical inference for human connectomes - PubMed ) which has a “virtual-lesion” method that enables you to prune and statistically validate the existence of an ROI-ROI tract against a whole-brain tractogram as a background. Nevertheless, this wasn’t computationally feasible for over 1000 ROI-ROI tracts and so we elected against it.
For 3) I am glad that you think that it should in theory work the same. We haven’t seen any papers out there use this approach (maybe we are too stuck in our little world and are missing out on other subfields of neuroscience implementing this), but its nice to hear someone else agree with us. @jdtournier@rsmith@ThijsDhollander Sorry to try and loop you guys in, but I wanted to see if you guys, as developers, see any issue with concatenating an ROI-ROI tract with a whole-brain tractogram prior to running SIFT, just in case we have a reviewer who may have a problem with this approach down the line.
