comparing fMRIPrep and HCP Pipelines: full-brain GLM images

As described in the introduction to this series of posts, the primary goal of these analyses is to compare the GLM estimates: is there a difference in the high vs. low cognitive control statistics produced from the different preprocessing pipelines (HCP, fMRIPrep) and spaces (volume, surface)? As explained in the introductory post, I've found comparing the statistics within Schaefer parcels most useful, since these are defined in all four results spaces.

But the parcel-averaged statistics were calculated from mass-univariate (every voxel and vertex separately) GLMs, and it's dangerous to make parcel averages without first looking at some of the full-brain GLM results to make sure everything seems sensible. So, here I'll show a few single-subject GLM results, before moving on to group and parcel-averaged in the next post(s).

The "buttons" GLMs are not of direct interest - they only model the button presses, not the cognitive tasks. But they're useful for a positive control: activity should be in motor areas and possibly visual (given the linking of response and visual change in some of our tasks). Below are the "button1" F statistics from the Buttons GLM for the same participant used in the previous posts. (Since these are TENT models, the Fstat provides a single summary of activity across the set of knots.)

First, the volume results. Since they're both in the same output space I can subtract the statistics voxel-wise, for the third row of images for each task. I subtracted HCP-fMRIPrep, so cool colors are voxels where the fMRIPrep-volume-preprocessed statistic was larger than the HCP-volume-preprocessed. As expected, the peaks are generally in grey matter, and pretty similar between the two pipelines: the two AX-CPT images look more similar to each other than do the HCP-preprocessed AX-CPT with the HCP-preprocessed Cued task-switching image. This is very good: I expect some variation between the three tasks (though all should have some motor), but the two pipelines ought to make similar statistics, given that they started with the same DICOMs, event timings, etc. While the two pipelines' images are visually similar, the difference images have quite a bit of blue in sensible brain areas, indicating that the fMRIPrep-preprocessed images had larger Fs in those voxels.

And here are the same statistics for the HCP and fMRIPrep-preprocessed surfaces. I couldn't subtract the two surfaces, since the two pipelines use different surface targets (so there isn't vertex-to-vertex correspondence; there are about 3x as many vertices in the HCP surface than fsaverage5). The inflation of the two underlays doesn't match exactly, but it's possible to tell that the statistics sort much more by task than preprocessing, which, as noted before, is very reassuring. There's also clear motor activity in all images, less in AX-CPT than the others, consistent with what is seen in the volume images for this person.

This person is quite typical, as are the conclusions drawn from this button1 F statistic: the full-brain statistical images tend to sort much more by statistic (which contrast, which task) than by preprocessing. But there are differences in the statistic magnitudes and cluster boundaries, and these are hard to evaluate qualitatively, particularly on the surfaces. Averaging the statistics within parcels makes it possible to quantify these differences, and so directly compare the four sets of results.