Scope

[tsalo]

I'd like to start running open multi-echo datasets through fMRIPrep and afni_proc.py, if an AFNI person could provide some recommendations there (@handwerkerd?), and publishing the results to GitHub and GIN. As long as I can get the process fairly simplified, I can run these jobs on UPenn's cluster.

@jsheunis's multi-echo-super repo (https://github.com/jsheunis/multi-echo-super) is very similar in scope, and I don't want to duplicate effort, so we can maybe figure out how best to divide any efforts. @jsheunis, I don't know if you're still working on that repo.

[tsalo]

Are there other pipelines (e.g., MRIQC) that would be good to run? Certainly including some things that are dataset-specific (e.g., phys2denoise) would also be good.

[jsheunis]

@tsalo happy to work together on this. I think it remains a good idea to make datasets and derivatives available as datalad datasets, but this does not have to be a blocker. If we continue with the datalad setup, it's perhaps a good idea to move that repo to tedana / me-ica organization? I'll make you admin on that repo.

[handwerkerd]

I'm not sure I'm completely following the whole plan, but if you ask for an afni_proc statement for dataset X that gives Y outputs, I can code up a template.

[tsalo]

That would be amazing, thank you! I'll follow up in #29.

[tsalo]

@handwerkerd do you think it would be a good idea to run NORDIC first on these datasets, even when phase information isn't available? I figure there are six options:

1. No NORDIC
2. Run NORDIC and replace the raw data/feed in only the NORDIC-denoised data into the preprocessing pipelines.
3. Run NORDIC, but add the denoised files as rec-nordic versions of the files to the dataset, so we get out NORDIC+preprocessing and raw+preprocessing derivatives.
4. - 6. Any of the above, but only do it for datasets with phase images.

[handwerkerd]

I would not run NORDIC. NORDIC would run on each echo time series separately and, given the method is still being tweaked and is a bit sensitive to the parameters of the data being used, I think it's very plausible that NORDIC will affect the relationship between echoes. It's fairly clear that NORDIC isn't perfectly removing only thermal noise and is adding some structural artifacts.

This is all addressable through improvements to NORDIC and better testing, but, at this point, I don't feel comfortable serially running NORDIC then tedana.

That said, I think the real solution is a combined method. NORDIC is looking for characteristics of Gaussian noise. When we have 3 echoes for each excitation pulse that should all have similar noise properties, the multi-echo informaiton might make the general NORDIC approach better.

[tsalo]

Thanks! That's a great point, and I really wish NORDIC leveraged multi-echo info.