Longitudinal FBA issue

Hi experts,

I am doing a longitudinal (2 time points) FBA following this pipeline (Fibre density and cross-section) and Genc et al. 2018.

Using TBSS, I found decreased FA from TP1 to TP2, which are also the common findings in the literatures. However, using the same data and FBA, I found the opposites, i.e. significant increased FD and logFC in some tracts.

My question is which step I could have done wrong in FBA? For my data, I have low b-value and single shell (b=1000 and 0). I used single-shell 3-tissue CSD to model the FOD. The inter-subjects FOD template and mask look fine. The number of fixels is also quite reasonable. For the statistical analysis, I calculated the changes TP1 - TP2, and used the one-sample t-test. The matrix includes 1s and 2 covariates:

1 2 1.2
1 1 0.5
1 1 -2.1

The contrasts are 1 0 0 for decrease and -1 0 0 for increase.

I noticed a reply in another topic:

How could I found the massive false temporal changes? Is there any other confounders that I am not aware of?

Thanks in advance!

Dear @XL258W ,

Did you use the same (averaged) response functions for both time points? Perhaps this might contribute to the result.


Hi Nick,

Thank you for the suggestion! But, I indeed used the same response function for all the subjects and both time points.

Hi @XL258W,

There’s a huge number of factors that vary in between the two analyses, so you would need to consider the full gamut of those differences in trying to isolate the primary source of the change in outcome.

The main one is the distinction between the quantitative metrics of interest, i.e. FA v.s. FD / FC. For instance, if you have a voxel containing two fixels, with one fixel having a greater value of FD than the other, but the underlying effect is that the value of FD increases in the smaller of the two fixels while the larger one is unaffected, then FA will decrease. So you would want to check for the prevalence of crossing fibres in locations where significant changes in both FA and FD have been observed. FC is a substantially different metric, and while there may be some correlation in observations between FA and FC, this could be due to e.g. greater or lesser amounts of partial volume affecting FA due to bundles being larger or smaller in cross-section.

Obviously the process of statistical enhancement is also massively different. One way you can interrogate these differences is to perform additional mix-and-match experiments; e.g. project FD and FC to voxels and test them using TBSS, and map FA to fixels and perform CFE.

I don’t personally have the experience with SS3T to comment on what it may or may not contribute to your observation, but from a purely experimental perspective you could try repeating the FBA using MSMT with CSF and WM responses only; that would hopefully confirm that that specific method has not introduced counter-intuitive results.

The low b-value is also worth considering carefully. The interpretation of the calculations performed to quantify the FD metric as being proportional to “fibre density” is predicated on both attenuation of the extra-cellular signal and no substantial variation in T2. It’s possible that for the actual process being studied here (combined with your acquisition), you do in fact have a decrease in underlying fibre density that contributed to the observed decrease in FA, but you also have a non-negligible extra-cellular signal with a long T2, which results in an elevated DWI signal magnitude and hence an increase in FD. So e.g. you could consider warping for each subject the mean b=0 and mean b=1000 image intensities to the template and comparing those data across the two time points. If you have non-DWI T2-weighted data (or better, T2 estimation), looking at differences there could be informative also.

How could I found the massive false temporal changes?

What you’re looking for primarily is fixels in which the value of FD is zero at one time point but not the other (other complexities are possible, but this is the easiest one). If your observations are in major white matter bundles this is less likely to be consequential though.

What you then have the option of doing is setting the value of the change in FD between the two time points to NaN. This will instruct the GLM to ignore the data for that subject in that particular fixel.

Is there any other confounders that I am not aware of?

Every single step in your whole pipeline introduces potential confounders. That exhaustive list is way too long for me to attempt to include here. But as far as your particular observation is concerned, I think the main thing to focus on is the limitation on the interpretation of the FD metric as being proportional to intra-axonal volume. If you think about it instead as capturing a change in the diffusion-weighted signal magnitude relative to fibre orientation, your result can probably be explained with some confidence.


1 Like