Workflow strategy

[tsalo]

@bbfrederick I think we can discuss the overall structure of the workflow here.

I'd like to keep the scope of each fMRIPost app fairly limited, so I want to stick to the rapidtide workflow, rather than any of the other CLIs in the rapidtide package (e.g., happy).

Are there any particular elements that rapidtide could/should use from preprocessing derivatives?
I have a few ideas:

• fMRIPrep confounds
• Tissue type masks?
• Physio regressors? From physiopy/prep4phys maybe.
• Phase data? From fMRIPost-phase

Can rapidtide leverage any of those? Are there any other features that it can/should use?

[bbfrederick]

I've recently been working on tweaking the analyses, and have found that using a gray matter mask to select the initial regressor, and also limiting the refinement and offset calculation to gray matter voxels, seem to stabilize the fit somewhat, so tissue masks would certainly be an option.

Also, motion (and confound) regression prior to starting analysis seems to help, so having access to the motion time courses from the confounds file would be good. Do the DVARS and FD time courses end up in the confounds file? Adding those to the confound regression could potentially also make the regressor extraction more robust.

ATM, I don't use physio regressors, although they could also potentially go into the confound regression. I'm not sure what I'd do with phase data.

[tsalo]

Thanks!

Do the DVARS and FD time courses end up in the confounds file?

fMRIPrep does produce DVARS and FD time courses in the confounds file. We might also want to consider filtering the motion parameters to remove pseudomotion, as recommended in Fair et al. (2020) and Gratton et al. (2020). I have the code to do that in XCP-D and can port it over.

I don't use physio regressors, although they could also potentially go into the confound regression

I assumed that at least some physio regressors would be similar to the sLFO signal at a given lag. Would regressing them out cause any problems with estimating the sLFO?

I'm not sure what I'd do with phase data.

I was thinking of maybe using the phase data to isolate macrovascular signal. Maybe rapidtide could run PCA/ICA on the complex-valued data?

[bbfrederick]

Thanks!

You're welcome!

Do the DVARS and FD time courses end up in the confounds file?

fMRIPrep does produce DVARS and FD time courses in the confounds file. We might also want to consider filtering the motion parameters to remove pseudomotion, as recommended in Fair et al. (2020) and Gratton et al. (2020). I have the code to do that in XCP-D and can port it over.

Rapidtide has a built in confound regression routine that's pretty flexible, but I'm open to modifying it if there's better code available.

I don't use physio regressors, although they could also potentially go into the confound regression

I assumed that at least some physio regressors would be similar to the sLFO signal at a given lag. Would regressing them out cause any problems with estimating the sLFO?

Maybe? Probably best to be careful about that. Straight up respiration can be a problem if it wanders into the LFO band (as it does a lot in older subjects) so having a way to effectively deal with it would be good. Rapidtide has a number of experimental options to try to deal with that, but I'm not super happy with any of them, so any progress on that front would be great.

I'm not sure what I'd do with phase data.

I was thinking of maybe using the phase data to isolate macrovascular signal. Maybe rapidtide could run PCA/ICA on the complex-valued data?

Sounds interesting... do you have a favorite reference for the macrovascular signal stuff so I can read up on it?

[tsalo]

Sounds interesting... do you have a favorite reference for the macrovascular signal stuff so I can read up on it?

This one might be good: https://doi.org/10.1016/j.neuroimage.2023.120011. Phase regression is a denoising method (also see nipreps/fmripost-phase#2) that downweights BOLD signal from macrovasculature and upweights BOLD signal from microvasculature by using fluctuations in the phase signal. I think rapidtide could theoretically use it in the opposite manner to isolate macrovascular signal, if that would be useful.

There's also a new spatial derivative of the phase data that could be used for something similar without the extra processing that goes into typical phase regression pipelines. See https://neurostars.org/t/applying-ln2-phase-jolt-to-multi-echo-data/30081.

[bbfrederick]

There's also a new spatial derivative of the phase data that could be used for something similar without the extra processing that goes into typical phase regression pipelines. See https://neurostars.org/t/applying-ln2-phase-jolt-to-multi-echo-data/30081.

Oh yeah - I talked to that guy at OHBM this year. That technique is super cool, and I was looking for a way to use it!