How can I apply a motion correction model based on 2018 fMRIPrep to 2022 fMRIPrep output?

I have a motion correction model trained based on a 2018 version of fMRIPrep. I’m now upgrading my pipeline to the 2022 version. The format of motion correction files generated in the 2022 version differ from the 2018 version, and I’m trying to work out whether it’s possible to avoid going to retrain my motion correction algorithm (which would be substantially more work), and feed it in data generate by the 2022 version. There are two factors to consider:

(1) The columns available in the motion correction tsv for 2018 fMRIPrep are a little different compared to 2022 fMRIPrep. I’ll explain below.

(2) Even if we can adapt the names of the columns, this would assume the correction algorithm is comparable, and that a motion correction script trained on the motion files from 2018 fMRIprep will behave as expected for 2022 fMRIPrep.

For dealing with (1): There are a number of column variants between fMRIPrep 2018 and 2022, but the main problem I’m having is that fMRIPrep 2018 had columns non-stdDVARS and vx-wisestdDVARS in addition to stdDVARS. fMRIPrep2022 has std_dvars, which I assume is equivalent to stdDVARS. However, the relationship between the old non-stdDVARS and vx-wisestdDVARS columns, and the new dvars and rmsd columns, if any, is unclear.

Other relevant differences are that 2022 fMRIPrep contains far more comp_cor columns, and it also contains derivatives and powers of overall X, Y, and Z translation scores.

Then, re (2) above, considering whether, even if I matched columns, there are any algorithms for 2018 and 2022 versions, and whether those differences would invalidate a motion correction model designed based on 2018 output–I don’t know where to start, though perhaps an empirical examination of the outputs of old and new versions would help. The different numbers of comp_cor columns output seems like a bad sign.

related: