Spurious group differences localized to the corticospinal tract in healthy control comparison: Protocol differences and CFE behavior at single-shell b=1000

1

Hi MRtrix3 team and community,

I am seeking advice regarding an unexpected finding during a Fixel-Based Analysis (FBA) comparing two cohorts of healthy controls.

Context & Data

We compared two independent cohorts of healthy controls acquired with slightly different scanning timeframes/protocols on the same general setup:

  • Control Group A: 24 healthy subjects

  • Control Group B: 14 healthy subjects

  • DWI Acquisition: Imaging was performed using a 3.0-Tesla MRI scanner (Ingenia, Philips) with a single-shot EPI sequence at a b-value of 1000 s/mm² (128 diffusion directions). Single-shell b= 1000 s/mm2 with 128 diffusion directions. (Preprocessed with standard denoising, degibbs, eddy/motion correction, and bias field correction).

  • FBA Pipeline: Standard Recommended Pipeline (single-tissue CSD for single-shell data, population template generation, non-linear registration, fixel cross-section and density computation, statistical inference via fixelcfestats with CFE and 5,000 permutations, covarying for age and sex).

Phenomenon

Despite comparing two healthy control cohorts (where we expected no significant structural differences), we observed significant group differences (pFWE < 0.05) specifically along the corticospinal tract (CST) across FD, FC, and FDC (Figure 1, 2, and 3). Rather than a global, diffuse baseline shift across the whole white matter, the “batch/protocol effect” appeared concentrated almost exclusively in the CST.

image
image2048×1330 140 KB
Figure 1. FDC

image
image2048×1330 141 KB
Figure 2. log_FC

image
image2048×1330 137 KB
Figure 3. log_FD

Questions

We understand that MRtrix3 strongly recommends multi-shell or higher b-values (b > 2500 s/mm2, directions >45) for optimal CSD resolution and elimination of isotropic free water, and that single-shell b=1000 data has known constraints.

Given this context:

  1. Biological vs. Algorithmic Localization: Why would scanner/protocol drift or batch effects manifest so selectively in the corticospinal tract rather than globally? Could the combination of high directional packing (single, prominent fiber population aligned along the z-axis) and gradient non-linearity/eddy currents make the CST particularly vulnerable?

  2. CFE Enhancement: Could the Connectivity-based Fixel Enhancement (CFE) algorithm inadvertently amplify small, continuous systematic drifts along long, coherent tracts into statistically significant FWE-corrected clusters?

  3. Handling Batch Effects in FBA: In the literature, ComBat is widely used for scalar DTI, but how does the community currently recommend harmonizing or controlling for subtle protocol differences when running FBA on fixel-wise metrics (FD/FC/FDC)?

Any insights, technical perspectives, or relevant literature pointers would be greatly appreciated.

Best regards,

2

Hi @KazuoWashida,

I’m not sure I understand what the differences were between the two control groups? You hint that there were slight differences in scanning timeframes / protocols – it would be good to have a bit more information on that front.

Otherwise, there is a chance the effect could be driven by the CFE, which will enhance ‘connected’ fixels. Since the CST is a large tract, it’s possible that it’s the most sensitive to a systemic effect than any other tract.

One way to check would be to look at the absolute effect size without thresholding by the FWE (1-p) map. If you find the effect is relatively uniform across the WM, but only shows significance within the CST, then that would confirm that you do indeed have a subtle systemic difference between the two cohorts – and that the CST is indeed most sensitive to these changes, presumably due to its particular shape and size.

If that’s not the issue, and the absolute effect is indeed primarily confined to the CST, there is a chance that you’re seeing a real effect – it depends on the specifics of what the difference is between the cohorts. The CST has a longer T2 than other white matter tracts, for reasons that haven’t (to my knowledge) been figured out (I personally suspect the larger axons in the CST mean less interaction of the intra-axonal water with the axonal membrane, and hence less scope for surface relaxation – but I have no evidence to back this up…). So if your cohorts were acquired with different echo times, it’s possible that the CST genuinely shows up as different…

That’s all speculation, but hopefully some of this might be useful…
All the best,

Donald

PS: apologies for the over-zealous spambot…