Hello MRtrix3 community,
I first wanted to say it has been a pleasure working with MRtrix3! I would like to express my sincere appreciation for all the hard work of the developers and the support they provide here and in the documentation. This, along with publishing so much of your work, is a huge service to the field. Thank you!
The goal of our current study was to map the axonal projections from two distinct regions in the ventral occipitotemporal cortex, to understand what areas these regions are structurally connected to (at the group level) and, importantly, the extent to which their projection targets differ within subjects.
I have a few questions that pertain to endpoint track density image (TDI) maps generated from the -ends_only option in tckmap. I have created these images for each ROI/subject in MNI152 space, then smoothed them using a 4mm gaussian kernel with mrfilter. I then performed a paired t-test and cluster correction using AFNI tools. The result shows large clusters of statistically significant difference in endpoint density, but the effect size across areas is quite variable. Before interpreting these contrast maps, I am interested in thresholding the result such that I eliminate projection targets which are not biologically “plausible” or “meaningful.” For instance, it seems that some streamlines reliably arrive in distant prefrontal areas from one ROI and not the other, but the average (SIFT2-weighted) endpoint density values there are very small, e.g. 0.002 in prefrontal vs 0.2 to posterior occipital areas. Note that this is exacerbated by the intentional over-sampling of streamlines projecting from these ROIs (see pipeline details below). A few questions:
- Can the resulting endpoint density values in each voxel be interpreted as indicating the cross-sectional area of arriving fibers? And, can a particular unit be ascribed to these values? (edited for clarity)
- Do you have any thoughts on whether/best way to threshold these maps?
- Does gaussian smoothing of TDI endpoint maps seem appropriate?
To briefly summarize my processing pipeline if it matters for the consideration of my questions (or in case it is helpful for others): my pipeline for this single-shell (b=2000) 3T data involved FSL-based preprocessing (including dwidenoise and distortion correction using Synb0-DisCo), Dhollander response function estimation, SS3T-CSD (using the the group mean response parameters), mtnormalize correction, tractography using iFOD2 with ACT / GMWMI seeding (based on FreeSurfer segmentation) / 10 million streamlines, and finally SIFT2. Prior to SIFT2, I generated an additional 500k streamlines from each ROI mask and combined these with the whole-brain streamline set (to more densely sample the projections from these ROIs). The ROIs were defined on the surface and sampled to volumetric space, respecting the cortical GM definition from FreeSurfer. Tractography was performed in subject’s native DWI space (after distortion correction), then a nonlinear transformation of the .tck files to MNI152 space was performed prior to TDI map creation/smoothing.
P.S. I look forward to the long-awaited paper by @rsmith which might help with some of these questions, though no pressure of course! 
Thanks again for your time.
-Ben
). I implemented the use of group-average response functions as well as intensity normalization via mtnormalize, which I believe satisfy the requirements laid out in your preprint regarding connection density normalization. I then apply the SIFT2 weights when performing tckmap to create the TDI/endpoint images (via -tck_weights_in flag). All of this together should amount to the appropriate scaling you mention here, but please correct me if I am wrong.