Ridding fMRI data of motion-related influences: removal of signals with distinct spatial and physical bases in multi-echo data
Jonathan Power, Mark Plitt, Stephen Gotts, Prantik Kundu, Valerie Voon, Peter Bandettini, Alex Martin
PNAS 2018 Feb 27; 115(9):E2105-2114
GODEC and ME-ICA code (Kundu Bitbucket)
Robust PCA code (.zip) and link to more implementations and an excellent tutorial
Below are movies associated with the article "Ridding fMRI data of motion-related influences: removal of signals with distinct spatial and physical bases in multi-echo data". All movies are 1080p and will look best at full-screen resolutions. The movies stream from YouTube and are organized as playlists. You can select cohort or subject from the playlist drop down menu at upper left of the movies (the three horizontal bars). Click on links next to the captions to download the movies. Files to download are hosted on Dropbox Pro; if the links don't work our traffic has exceeded its 200GB/day limit and the links will be re-enabled the following day.
Video 1 (.mov - 150MB):
This video shows ME-ICA and FIT procedures separating S0 from R2* effects in the ME cohort, for the 87 subjects who successfully completed both procedures. A second frame per subject shows the mean signal in the cortical ribbon for the undenoised and denoised data after z-scoring all traces (post-denoising magnitudes are substantially smaller, as expected, hence the z-scoring). Then several frames show the time series of all ME-ICA components and how they correlate with each other and the mean cortical signal.
Video 2 (.mov - 10MB):
This video shows the link between respiration and BOLD signals in NA data. At top, motion traces, at middle, heart rate and respiration traces, and at bottom, in two frames, the FIT R2* and ME-ICA denoised BOLD signals. Heart rate and respiration are obtained from independent traces. The cyclic modulation of heart rate by respiratory cycle is due to chest in intrathoracic pressure, as is the transient elevation and then decrease of heart rate upon very deep breaths.
Video 3 (.mov - 55MB):
This video shows that global signals aren't removed much by common nuisance regressions. The left column of grayscale panels shows ME-ICA denoised data at top, then at middle the same data after regressing 12 motion parameters and mean white matter and ventricle signals, and then at bottom the same regression but also including the mean cortical signal. The right columns show the same procedures, just starting from raw data at TE2. Note that the nuisance regressions use signals from the deepest white matter and ventricle compartments, so that gray matter signal isn't inadvertently removed by partial volume effects (superficial voxels in the nuisance compartments, adjacent to gray matter, have signals almost identical to gray matter signals).
Video 4 (.mov - 40MB):
This video shows 2 methods to remove respiratory signals from fMRI time series. At left, the separation of ME-ICA denoised data into low-rank and sparse signals by GODEC. At right, the variance removed and retained by regression of mean gray matter signals in the same data.
Video 5 (.movs ~1GB per link): ME01-15 ME16-30 ME31-45 ME46-60 ME61-75 ME76-89
This video shows Workbench surface renderings of signals in "optimally combined", ME-ICA discarded, and ME-ICA retained images, as well as the variance discarded and retained after GODEC and mean cortical signal regression. Scaling of the heatmaps is held constant across images. At top position and motion estimates are shown, as well as the signals at the 264 ROIs used to generate correlation matrices of the paper. Automated image generation from Workbench crashes not infrequently, and only 73 subjects' data are shown as a result of sporadic crashes (each movie requires 239 volumes x 2 surfaces x 7 images = 3346 correct calls to image generation).
Video 6 (.mov - 45MB)
This video shows gray plots of ME subjects undergoing 4 techniques to remove global signals from fMRI data: robust PCA (RPCA), Go Decomposition (GODEC), global signal regression (GSR), and CompCor. At top are the ME-ICA denoised data, and then the two plots below show the global variance and the sparse variance identified by each method. RPCA failed on some subjects, hence the missing images.
ADDENDUM 18 July 2018:
Reply to Spreng et al.: Multi-echo fMRI denoising does not remove global motion-associated respiratory signals
Jonathan Power, Alex Martin
Comment and reply (.pdf)
** Spreng et al. wrote a letter to PNAS questioning the basic findings of our paper. We were asked to respond, and wrote a reply, after which PNAS decided not to publish the letter and response. We post the reply here as it contains some new information that may be of use to readers **
Spreng et al. wrote a letter to the PNAS editors criticizing our findings and stating that they do not find a relationship between respiratory variability and whole-brain fMRI signal variance. We responded with a figure demonstrating the relationship between respiratory variability and whole-brain signal variance in both runs of the NA data separately and also for the NIH subjects of Power et al, 2017. Spreng et al. also suggested that ME-ICA can remove respiratory signals from a scan. It does not, for reasons outlined in the paper. We illustrated this failure to remove respiratory signals with examples of whole-brain signals caused by isolated deep breaths in several subjects existing before and after ME-ICA.
The data of all of these subjects were visually inspected by myself as part of their selection for analysis. I suggested several reasons that Spreng et al. may have not detected the influence of respiration in their data. Two of these reasons have to do with checking the quality of respiratory traces and calculating the envelope of these traces, and a third has to do with looking at scans to see that there are not other prevalent global artifacts (especially scanner problems or motion) that would obscure relationships of respiration and global variance. The video below illustrates how to check such properties, using the NA scans.
To calculate the shown envelopes in Matlab (for a 400Hz respiratory trace):
% variability of respiratory envelope
z = zscore(respbelt);
[yu yl] = envelope(z,4000,'rms');
respvar = std(yu);
Video for Spreng et al. (.mov - 5 MB)
The video below shows the respiratory traces (in blue) and respiratory envelopes (in orange) for each of the 2 runs for each of the 12 NA subjects. Additionally a gray plot of the timeseries at 264 regions of interest from Power et al., 2011 (those used for motion-dependent analyses) are shown to convey scan properties. Note that the physiologic records often last beyond the scanning period. In the paper, we inadvertently included the entire physiologic record in calculations used for Figure 2, and in response to Spreng and colleagues, we re-analyzed the data and discovered the mistake and re-present the relevant analyses with the corrected statistic. Our corrected statistics strengthen the link between respiratory variability and global fMRI signals beyond what we initially reported. The entire physiological trace is shown at the top of this video, and the scan-limited trace is shown in the middle panel, along with summary statistics (std of respiratory envelope). Additionally, we indicate whether the scan passed all versions of processing used in the paper and whether it was included in the analyses of Figure 2 of the main paper. The relevant figure is reproduced below.
The revised analysis of Figure 2 with pertinent additional data is shown below.