Dear all,

Is there a way to regress against the global signal in CONN?

Thank you!

Best,
Natalia

Dear Natalia,

By default the conn toolbox uses CompCor, instead of global signal regression, to address potential subject-movement and physiological confounding effects without the risk of artificially introducing anticorrelations into your functional connectivity estimates, so generally I would not recommend using global signal regression instead. See this thread (https://www.nitrc.org/forum/forum.php?th...), or this reference (Chai X.J., Nieto-Castanon A., Ongur D., Whitfield-Gabrieli S. (2012) Anticorrelations in resting state networks without global signal regression. NeuroImage 59(2):1420-1428) for more details about this issue.

That said, if you still want to regress out the global signal you may do so (approximately) by entering the 'grey-matter' ROI into the 'confounds' list in the 'Preprocessing' tab (and probably removing the White-matter and CSF ROIs if you do not want to use the CompCor strategy), or (perhaps more precisely) by adding a new ROI in the 'Setup->ROIs' tab that encompass the entire brain (e.g. brainmask.nii if you are working on normalized volumes) and entering this ROI instead into the 'confounds' list in the 'Preprocessing' tab. 

Hope this helps
Alfonso
Originally posted by Natalia Yakunina:

Thank you yet again for the clear answer!

Sincerely,
Natalia

Alfonso,
 
What if one were conducting a pre-/post- experiment in which global cerebral perfusion may be reasonably assumed to increase between baseline and follow-up?  Would aCompCor still correct for this sort of global increase in components feeding the BOLD response? 
 
Thanks,
Jeff

Hi Jeff,

That is a very interesting question. As far as I know (but I am not an expert on this topic so please take the following comments with a grain of salt) cerebral perfusion effects on the BOLD signal are, to a first-order approximation, modeled as additive (changes in the BOLD signal baseline) and/or multiplicative (changes in the BOLD signal scale or sensitivity to vascular and physiological sources) effects. For neither of these effects you would expect to need any additional correction in functional connectivity analyses, since: a) global additive differences between sessions in the BOLD signal are controlled by session-specific regressors and high-pass filtering steps -which remove the mean BOLD signal separately for each session- and by the mean-invariance property of the correlation measures -which also disregards mean BOLD signal levels-; and b) scaling/multiplicative differences between sessions in the BOLD signal are controlled by global signal scaling and the scaling-invariance property of the correlation measures -which disregard the actual scale/units of the BOLD signal- combined with the linear nature of all of the denoising steps -re-scaling the entire BOLD signal results in simply a re-scaled version of the denoised BOLD signal-. Of course other, more complex, effects may be at play beyond global BOLD signal baseline/scaling differences between the sessions, but I would behard-pressed to imagine an scenario where global signal regression would do a better job at correcting for these effects than aCompCor (but please let me know your thoughts or any additional details about possible perfusion-related differences that you may expect and I will be happy to comment specifically on how to best correct for those effects)

Best
Alfonso

Originally posted by Jeff Browndyke:

Thanks for the (as usual) wonderful response, Alfonso.  I asked because this came up as a criticism by two separate reviewers of a paper we recently submitted.  Coincident with the data we collected we also obtained ASL, which did show a global increase in perfusion between baseline and follow-up in one of our groups, but this increase was only found for summed global perfusion.  There were no regional perfusion increases that survived uncontrolled p<0.001.  To address the criticism, I was planning on running suggested non-aCompCor work around and see if there is any appreciable differences in the global signal correction vs. aCompCor approach.  Any particular issues I need to watch out for when setting up this non-aCompCor processing pipeline?

Warm regards,
Jeff

BTW - another criticism we weathered was the thought that our resting state sequence duration was too short (~6 minutes).  One of the reviewers insisted that we needed 10 minute runs, which in my mind just opens one up to problems with increased movement artifact (particularly in our older patient samples).  Do you know of any references off hand that indicate that ~6 minute runs are adequate?

Jeff

Hi Jeff,

Yes, you will find plenty of references indicating that 6 minutes scanning sessions result in moderate to strong reliability for resting-state functional connectivity measures. For example you may point to Shehzad et al. 2009. "The Resting Brain: Unconstrained yet Reliable" (this is also the reference for the NYU test-retest dataset, in our own Whitfield-Gabrieli & Nieto-Castanon, 2012 manuscript we used this same dataset to show strong group-level reliability of connectivity measures using CONN). Another very good manuscript in this regard is Van Dijk et al. 2010. "Intrinsic Functional Connectivity As a Tool For Human Connectomics: Theory, Properties, and Optimization"

The reviewer is perhaps thinking along the lines of Birn et al. 2013. "The effect of scan length on the reliability of resting-state fMRI connectivity estimates", which recommends longer (12minutes or above) scanning sessions and show how that consistently increases the reliability of subject-level connectivity measures. If that is the case, perhaps you can simply point out that increases in within- and between-session reliability are only expected to increase your analysis power, but not change their validity (i.e. any significant results you are showing are still expected to be replicated in a higher-power experiment). There are, of course, many other factors which affect and limit the power of your analyses or your design choices. As you suggest practical concerns such as the ability to keep your subjects with minimal motion and compliant is one important consideration when determining the scan length. Also it is important to keep in mind that the between-subjects variability in connectivity measures is typically very large, and this very strongly limits the impact of improvements in the reliability of your single-subject measures on your second-level analysis sensitivity/power (e.g. doubling the number of subjects will typically have a much larger impact on your analyses power than doubling the scanning length). 

Hope this helps
Alfonso
Originally posted by Jeff Browndyke:

Thanks, Alfonso! 
 
BTW - when I was re-running our data through pre-processing.  I took a look at the Matlab processing log.  There are mean and SD values listed in this log for each subject with respect to "mean global" and well as movement parameters?  Is this "mean global" the same as mean global signal?  If so, then maybe I could just compare these output values between sessions and groups to show that there are no significant differences and forego the whole non-aCompCor reanalysis?
 
Warm regards,
Jeff

Dear Alfonso,

when you say adding a new ROi that encompasses the whole brain would be a way to approximate global signal regression, would you remove the WM and CSF ROIs from the confounds list or not? If not, wouldn't that be redundant somehow?

Cheers,
Sascha

Originally posted by Alfonso Nieto-Castanon:

Dear Sascha,

I do not recommend using GSR because: a) it provides a less comprehensive correction compared to aCompCor (it does not really account for potential spatial differences in the way physiological effects are expressed in your data); and b) it can introduce artifactual anticorrelations in your results (see Murphy et al. 2009). If you really want to use GSR then, given that you are already deciding to live with the (b) issues above, I would probably recommend to still use aCompCor as well (keeping WM/CSF principal components) to at least address the (a) issues above. Yes, in that case the global signal will be partially redundant with the components extracted from the WM and CSF areas, but it will still introduce some new aspect to the regressed effects (unfortunately, that new aspect that is being introduced will be the portion of the global signal that is not already present in the WM and CSF components, which most likely represents "true" neural-source signal, so, again, perhaps not such a great idea to use GSR....)

Hope this helps
Alfonso
 
Originally posted by Sascha Froelich:

Thanks for further clarification on GSR concerns, Alfonso.

When using aCompCor, have there been documented recommendations for *how* to choose the explicit number of principal components to be regressed from WM and CSF regions?  The Chai et al. (2012) methods paper gives results with regression of 1, 3, 5, and 10 principal components per each WM and CSF regions and eventually settles on 5 for extensive analyses; however, the choice of 5 seems pretty arbitrary. There are bound to be differences in the PC structure of research subjects' data, and I worry that choosing an arbitrary cutoff would result in either insufficient removal of "noise" or distortion of "neural signal." The latter case I think would be increasingly likely when choosing aCompCor with large numbers (e.g., 5 per CSF and WM) of components removed (is anything >1 safe? if not, why not just use the average?). 

A method to determine on a single-subject basis the optimal number of artifact components to regress using aCompCor would be greatly appreciated. Are you aware if anyone out there in the CONN-world has implemented this?

Hello,

On the CONN website, it indicates that GSR can be completed by "simply defining a new ROI encompassing the entire brain (e.g. a subject-specific brain mask resulting from the outlier detection step during preprocessing, or gray matter mask resulting from the segmentation step) and using this ROI as an additional potential confounding effect in the standard Linear Regression denoising step."

If I include a subject-specific grey-matter mask in the Confounds section of the denoising step, should I still include the WM and CSF ROIs as Confounds? Or would the combination of CompCor + GSR become an issue?

Thank you,
Ben

 
Hello all,
I am writing a conn batch script which does global signal regression along with other confounds. I share a list of confound names:


But I am not able to understand how should I add the name of a whole brain mask in the list. Will it be "art_mask_aurest.nii" file? If someone knows the name of the whole brain mask file generated by CONN's default prepocessing pipeline, please let me know. I am assuming that the following with appropriate filename should work:-



Thank you and regards,
Rudradeep