comparing fMRIPrep and HCP Pipelines: GLM TENT curves

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)? The previous step of the analysis confirmed that the whole-brain GLMs produced sensible results, so here I'll start showing the statistics averaged within parcels. As noted before, averaging the statistics within Schaefer parcels is especially useful here, since they are defined in all three results spaces (MNI volume; surface fsaverage5, surface HCP). We've been using the 400-parcel resolution in our lab, so I adopted it here as well; I haven't tried with other resolutions.

Below is a detailed explanation of how to read these figures. They're complex, but the best way I've found (so far!) to sensibly compare sets of GLM results for this dataset. Surface-derived curves are shown in the top pair of figures (with the surface brains), and volume-derived in the lower (with the volume brains). These are group (over the 13 test people) estimates for two informative parcels (90 and 91 in the left hemisphere). Files with all parcels can be downloaded here; this zip contains the compiled pdfs (that I took these images from) as well as the source knitr (.rnw).

There are five panes in each row. The first has a picture of the parcel and its name, surface for surface statistics (though I need to fix the text label spacing) and volume for volume statistics. The bar graphs by the parcel image show the sustained (block) regressors (robust mean and SEM) for each of the four tasks, in alphabetical order (same as the line graphs). Aside: the volume bar graph has space for "P" and "R" - these are the proactive and reactive DMCC sessions. I adapted this knitr template from one that included all three sessions; please ignore the empty spaces and "PRO" "REA" labels (fMRIPrep preprocessing was only run on Baseline session images).

The other four panes show the event-related part of the model for each task; the two regressors subtracted to make the high-low cognitive control contrast for each task are listed in the title after the task name. In the curve figures, the thick line at each knot (x-axis) is the across-subjects (robust) mean coefficient difference; the thin lines are the (robust) SEMs. Trimming of 0.1 was used in both cases; trimse from the WRS2 package was used for the SEM and base R for the trimmed mean. The solid line is calculated from the HCP-preprocessed images, dashed with fMRIPrep (abbreviated fp). The small bars below each set of lines gives the difference at the corresponding knot; HCP-fMRIPrep (so if the HCP-derived mean coefficient is larger the difference will be positive and the little bar will be above the line, on the side marked with "hcp").

The number of knots (x-axis) in each graph varies, since the tasks vary in duration: Sternberg trials are longest, Stroop trials are shortest. Each knot is 2 TRs (2.4 ms) in all tasks. The Stroop curve looks a bit HRF-ish; since each trial is just a few seconds, as is reasonable. The others are a bit delayed or doubled (in the case of Sternberg) due to the trial dynamics; there's lots more detail on the DMCC website, if you're curious.

For the purpose of the preprocessing comparisons, it's most relevant that we identified one knot for each task that should be most relevant for these high vs. low cognitive control contrasts, which is shaded in grey on the plots. Focusing on a single knot in each task gives a way to summarize the curves - we're not equally interested in all knots' estimates - which will be shown in the next post. Note that which knots are "most relevant" and which contrasts (and GLMs) would be used for high vs. low was determined before these preprocessing comparisons were conceived, primarily on the basis of task design, but also checking against the volumetric HCP-preprocessed GLM results.

I described these two parcels (90 and 91) as "informative" because in all four cases (surface, volume, HCP, fMRIPrep) the TENT curves have a fairly sensible shape (given the task timing) in all tasks, with a positive (high > low cognitive control) mean at the target knot (shaded). These two parcels are also located in areas that are sensible for having more activity in high cognitive control conditions. In the next post I'll show whole-brain results again, but summarized within parcel and at the target knot.