comparing fMRIPrep and HCP Pipelines: motion regressors

See this post for the introduction to this series of posts. Here, I'll describe and compare the motion regressors produced by the fMRIPrep and HCP Pipelines. Short version: the file formats are a bit different, but the estimated motion is nearly identical.

Long version: the first step is figuring out where the motion regressors are put by the two pipelines. If you kept the HCP file structure, there will be one Movement_Regressors.txt file for each preprocessed run. The files are always called Movement_Regressors.txt; you track which one goes with which run and person by the directory structure. For example, the file for subject 203418, run SternBas2_PA is /MINIMALLY_PREPROCESSED/203418/MNINonLinear/Results/tfMRI_SternBas2_PA/Movement_Regressors.txt.

The above image is the first bit of the file, read into R. It has 12 columns: x, y, z, roll, pitch, yaw, then their six derivatives (stated on page 75 of HCP_S900_Release_Reference_Manual.pdf). The rotation columns are in degrees, translation in mm. There are 600 rows in this particular text file, matching the number of volumes acquired in the run.


fMRIPrep outputs a file ending in _confounds.tsv for each run, with its location and name following the BIDS naming conventions. For example, the motion regressors file for the same subject and run as above is /sub-203418/fMRIPrep_output/fmriprep/sub-203418/ses-01/func/sub-203418_ses-01_task-Stern_acq-MB42p4PA_run-02_bold_confounds.tsv.

The above is the first bit of the file, read into R. As you can tell in the screenshot, this file has many more columns than the HCP version: 35 instead of 12 (the same number of rows, though: one per acquired volume). The columns are named, and the documentation gives details on each. The motion regressors I want are in six columns, called X, Y, Z, RotX, RotY, and RotZ. "RotX","RotY","RotZ" correspond to roll, pitch, and yaw, but are in radians; translation is in mm.

UPDATE (24 April 2019): fMRIPrep version 1.3.2 names the columns "trans_x", "trans_y", "trans_z", "rot_x", "rot_y", "rot_z".

Now I can plot both sets of motion parameters to see if they vary. I converted the fMRIPrep rotations to degrees, then plotted the six estimates from each program in the top of this figure. (TRs (1.2 ms) are along the x-axis; the vertical grey lines give 1-minute intervals.) As you hopefully can see (click the figure to enlarge), it's actually hard to tell that there are 12 lines in the top - those from fMRIPrep and the HCP pipelines are so close that it mostly looks like there are only 6 lines.

The lower frame has the FD version of the same two sets of motion estimates; HCP Pipelines in black and fMRIPrep in green (this is unfiltered FD, and our censoring threshold of 0.9 is shown in red). A little separation between the two FD lines is visible, but again, quite minimal.

I picked this run for this post because there's a bit of actual head motion, and if you zoom in around frame 525 you can see that there is a little separation between the translation estimates. I had to check a few runs to find one with any visible separation; in our 13 test people (with 8 runs each) this is pretty much as large as the discrepancy gets, which is why at this time I'm describing the motion estimates from fMRIPrep and the HCP pipeline as essentially identical.

Note: The figure above is from a knitr template that I made for doing these QC comparisons; see note at the end of this post.