comparing fMRIPrep and HCP Pipelines: tSNR images

See this for the introduction to this series of posts. Yesterday I showed that the motion regressors produced by the two programs are nearly identical (as we'd hope, since they started with the same DICOMs!). But how do the preprocessed images compare? Short version: the preprocessed images are qualitatively similar in appearance (and standard QC images), with differences between people more obvious than differences between the pipeline. But only "similar", not "nearly identical" (which is how I described the motion regressors).

(Much) Longer version: The main goal of these preprocessing tests is to compare the GLM estimates (described a bit in the introduction and more later), not to perform QC on the images from the pipelines. This adds a complication, however, because the images that go into the GLMs are not exactly the same as the images that come out of the pipelines. It strikes me as most useful to look at the images as ready to go into the GLMs, since that's the final step at which things vary - identical GLMs were performed at every voxel and vertex.

We wanted to make the GLM estimate comparisons as direct and fair as possible, so aimed to do equivalent transformations of the surface and volume preprocessed images from each pipeline. If you think we missed that target, please let me know. We use afni for the after-preprocessing steps, as summarized here. First, both fMRIPrep and the HCP pipelines produce MNI-normalized volumes, but the bounding box is a bit different, so we used 3dresample to match the fMRIPrep images to the HCP, as well as to reorient to LPI. Both the fMRIPrep and HCP pipeline volumes are then smoothed (3dBlurToFWHM, -FWHM 4); surfaces are NOT smoothed with afni commands, since they were smoothed in the preprocessing pipelines. Finally, the timeseries at each voxel (or vertex) was scaled (in the usual way for afni, see this or this), using 3dcalc -a full.timeseries -b mean.timeseries -expr 'min(200, a/b*100)*step(a)*step(b)', where mean.timeseries is an image of the voxel (or vertex)-wise means, calculated with 3dTstat.

Single frames (this is #100 from the same run as in the previous post) from the ready-for-afni images look similar; an obvious difference is that the HCP pipelines mask the brain, while fMRIPrep does not. I have the color scaling the same in both images (90 to 100) and the intensity in the brain is similar (as it should be, given the scaling); the voxel under the crosshairs has the value 99.8 after fMRIPrep and 98.6 after the HCP pipelines.The frontal dropout (yellow arrow) is obvious in both, but looks like weird striping in the HCP image because of the masking; without masking it blends into the background. Spot-checking frames of the images I can't say that I've found one preprocessing consistently "better" than the other; by eye, sometimes one looks a bit better, sometimes the other.

Here's the same frame, from the surfaces. The HCP pipelines project to a different underlay surface anatomy than fMRIPrep, so the brain shapes are a bit different (fMRIPrep to fsaverage5, 10242 vertices/hemisphere; HCP with 32492 vertices/hemisphere). To be honest, I find these harder to interpret ... swirly colors.

If the single frames look similar, what about QC images? Well, by eye they also strike me as "different but similar", without an overwhelming preference for one being better than the other. The above are tSNR images of the same test run, all with the same color scaling.


Is the story in this whole series of comparison posts going to be "similar"? No, differences turn up in the GLM results, especially when summarizing within parcels ... stay tuned.

Note: the movement graphs and tSNR images were made with R knitr code. The full pdf (four tasks for this person) is here, and its source here. The paths/etc. won't work to compile the knitr (unless you're on our network!), but the plotting code and afni commands may prove useful; the sourced gifti plotting functions are here.