I’ve searched the archives and found a couple of related topics, but nothing that addressed this question directly.
We have acquired two sets of DWI data on our participants, one with multiple b-values (max = 2000, TE = 88 ms) and a second with b = 3000, but with a longer TE (100 ms). Each acquisition has it’s own set of b=0 scans. We would like to combine the two acquisitions for fixel processing.
The older discussions I’ve seen on the forum seem to suggest that this might be done by applying a global scaling factor, based on the global signal differences between the two sets of b=0 scans. The problem with this approach is that this does not take into account differences in T2 across the brain, and it seems that a better approach would be to use the b=0 images (mean) to scale all of the b > 0 images, and feed these normalized images into the FOD calculations. After all, this is presumably what the DWI processing will apply to the data in order to estimate the effect of the diffusion weighting, so why can’t we apply it upfront (and leave out the b=0 images)?
I’m just not sure if the code can handle data without any b = 0 data, or if there are any other untold downstream effects that these normalizations will have on the processing stream.
In general, it’s considered a Bad Idea™ to mix echo times across shells… Some recent techniques look specifically at T2 & diffusion concurrently, but they require bespoke acquisitions, and will generally acquire the same b-values over different TEs to be able to disentangle all the different effects. That’s not the case with your data though.
However, as it turns out, MSMT-CSD is probably one of the few techniques out there that isn’t really impacted by different TEs being used for each shell. That’s because it doesn’t rely on a specific signal model (e.g. the diffusion tensor, or ball & sticks, etc.). Instead, it estimates the signal response within each shell for each tissue type directly from the data, and the effect of T2 differences is absorbed in these responses.
Slight disclaimer: this is based on theoretical considerations, I haven’t seen anyone actually investigate how valid this might be with real data.
So you should be fine to run MSMT-CSD for your analysis, provided:
all your datasets have the same combination of TE & b for each shell
you take appropriate steps to match the signal intensities across your two shells (see below)
you only run MSMT-CSD. I don’t think you could run DKI or NODDI on these data without inviting some criticism – though personally I doubt that the impact of the difference in TE would be enough to invalidate any observed group differences, etc. It might however impact on the accuracy and interpretation of these differences.
To perform the matching of signal intensities, you can use the dedicated dwicat command. This does the matching and concatenation based on histogram matching of the b=0 images. And yes, this is required, because:
This might indeed be sufficient for many other techniques, but that’s not how CSD or MSMT-CSD work. There is no normalisation to the b=0 signal, and for MSMT-CSD, the b=0 data are used in the fit. For CSD, it’s important that the raw DW intensities are scaled appropriately between volumes / shells.
Thanks so much, very helpful. Yes, I’m aware of the issue wrt DTI and NODDI-type fitting and this wasn’t our intention, it was specifically with respect to CSD.