Guidance on FBA pipeline for single-group ROI-to-ROI connectivity analysis

1

Hi everyone,

I’m new to diffusion analysis and the MRtrix3 ecosystem, and I would be incredibly grateful for a sanity check on an analysis plan. My apologies if any of these questions are naive.

1. Background & Initial Problem:

I have preprocessed 35 subjects from the MGH-USC HCP dataset. I’ve run dwi2response dhollander and dwi2fod msmt_csd to generate high-quality FODs for each subject.

My primary scientific goal is to characterize the structural connectivity between a seed region in the posterior insula and the various thalamic nuclei (defined by the Morel atlas).

My first approach was a standard seed-based connectome analysis. For each subject in their native space, I ran:

  • Anatomically-Constrained Tractography (tckgen) seeded from the seed file.

  • SIFT2 weighting (tcksift2) to get biologically plausible streamline weights.

  • Connectome generation (tck2connectome) to get a final matrix of connection strengths.

However, the resulting connectivity values showed very high inter-subject variability, making it difficult to draw firm conclusions about which connections are truly the most consistent and prominent.

2. Different approach: A Fixel-Based Analysis (FBA) Pipeline

To address the high variability, I am now planning to use FBA to get a more robust, spatially specific result. Since I only have a single group of healthy subjects and no behavioral data, my main uncertainty is whether the following pipeline is a valid way to answer my ROI-to-ROI question.

Here is my proposed step-by-step plan:

  • Step 1: Create a Population Template. Use population_template on all 35 of my wmfod_norm.mif files to create an unbiased, group-average FOD template.

  • Step 2: Subject-to-Template Registration. Use mrregister to compute the warp field that aligns each subject’s FOD to the new group template.

  • Step 3: Calculate Fixel Metrics. For each subject, upsample their FODs and warps to the template, and then run fod2fixel, fixelcorrespondence, and fixelcalc to generate the final FDC map for each subject in the template space.

  • Step 4: Whole-Brain Group Statistics. Run fixelcfestats to perform a whole-brain statistical analysis. My plan is to use a one-sample t-test (a design matrix with a single column of 1s and a contrast of [1]) to identify the “core” white matter skeleton of fixels that are significantly and consistently present across my healthy cohort.

  • Step 5: Create an Anatomical Tract Atlas. On the wmfod_template.mif, generate a large whole-brain tractogram (e.g., tckgen -seed_dynamic) and filter it with tcksift2 to create a high-quality anatomical map of all major pathways.

  • Step 6: Isolate Tracts of Interest. For each thalamic nucleus I want to investigate, I plan to use tckedit with my seed ROI and the specific nucleus ROI (e.g., -include seed.mif -include VPI_nucleus.mif) to select the streamlines connecting them from the whole-brain tractogram.

  • Step 7: Create Tract-Specific Fixel Masks. Use tract2fixel to convert each of the specific streamline files (e.g., insula-to-VPI.tck) into a binary fixel mask.

  • Step 8: Quantify Tracts and Compare Consistency. This is the final analysis step. I plan to use the fixel masks from Step 7 to:

  1. Extract Mean FDC: For each subject, and for each specific tract mask, use mrstats -mask on their FDC map.

  2. Analyze and Compare: With these values, I can now calculate the group-level descriptive statistics (mean, SD) for each pathway.

3. My Core Questions:

This is the part where I would most appreciate your guidance.

  1. Does this overall single-group FBA workflow make sense for my goal of characterizing and comparing the consistency of different ROI-to-ROI pathways?

  2. Is the one-sample t-test (design=[1;1;…], contrast=[1]) the correct statistical approach to define a meaningful “core” white matter skeleton in a healthy cohort?

  3. What is the best and most valid way to use the final tract-specific fixel masks (from Step 7) to get my final answer? Should I:

  • A) Simply use the masks to visualize which of the tracts were found to be consistent in the whole-brain analysis (from Step 4)?

  • B) Use mrstats -mask with my specific fixel masks on each of the 35 subject FDC maps to extract a mean FDC value per tract, per subject? This would give me a quantitative value that I could then use to compare the different connections.

  • C) Or is there a more standard or powerful method that I’m missing?

Thank you so much for your time and for developing these incredible tools. Any advice you could offer would be a huge help.

Best regards,

Govind

2

Hi Govind,

between a seed region in the posterior insula
…

  • Anatomically-Constrained Tractography (tckgen) seeded from the seed file.
  • SIFT2 weighting (tcksift2) to get biologically plausible streamline weights.

Please don’t run SIFT2 on exclusively targeted tracking data:

My plan is to use a one-sample t-test (a design matrix with a single column of 1s and a contrast of [1]) to identify the “core” white matter skeleton of fixels that are significantly and consistently present across my healthy cohort.
…

  1. Is the one-sample t-test (design=[1;1;…], contrast=[1]) the correct statistical approach to define a meaningful “core” white matter skeleton in a healthy cohort?

I can somewhat see the logic here, but I’m not sure if this is more obtuse than it needs to be:

  1. The test here doesn’t actually address “consistency”; it only evaluates whether the group mean is greater than zero. A small number of subjects with high-density fixels, and all other subjects with zero, could nevertheless achieve this.
  2. While not explicitly shown, the likely “default” would be to perform such a test downstream of fixel data smoothing, which will introduce non-zero values in template fixels with no corresponding subject fixel.
  3. It’s not clear why specifically CFE should be applied in classifying individual fixels as having good vs. poor consistency.

Just putting a threshold on the fraction of subjects that need to have a non-zero value might suffice & be more direct / reportable / interpretable.

  1. Does this overall single-group FBA workflow make sense for my goal of characterizing and comparing the consistency of different ROI-to-ROI pathways?
    …
  • B) Use mrstats -mask with my specific fixel masks on each of the 35 subject FDC maps to extract a mean FDC value per tract, per subject? This would give me a quantitative value that I could then use to compare the different connections.

Yes, this is the standard approach for more robust tract connectivity. FBC as a metric is more faithful to the biological property of interest, but subject-specific tractography is just not robust enough in a lot of use cases.

  • Step 7: Create Tract-Specific Fixel Masks. Use tract2fixel to convert each of the specific streamline files (e.g., insula-to-VPI.tck) into a binary fixel mask.

tck2fixel won’t implicitly give you a binary fixel mask. You would need to make an explicit decision of how to threshold the fixel-wise streamline density. It’s fine if that threshold is one streamline, just better to understand that that’s what is happening.

  • A) Simply use the masks to visualize which of the tracts were found to be consistent in the whole-brain analysis (from Step 4)?

I’m not sure exactly what’s being proposed here or why:

  1. If you are proposing going to the extent of a one-sample t-test to assess “consistency” of fixels in the cohort, then it’s unclear why you would want to then limit your investigation of these tracts of interest to merely “visualisation”.

  2. Be careful about language / interpretation here. Such an outcome would be aptly described as “the whole-brain analysis result was inclusive of this tract”, whereas the description “this tract was significant” would implicitly convey a test other than what was performed. Cluster-based statistical inference has come under some criticism in this respect. Just because some region is included in a significance mask doesn’t mean that there is an effect present in that region, only that it is connected to something where an effect is truly present. CFE is more robust to this than other cluster-based tests, but it nevertheless warrants care.

  • C) Or is there a more standard or powerful method that I’m missing?

Not yet. :smiling_face_with_horns:

Regards
Rob