help > How to denoise with global correlation GCOR
Showing 1-7 of 7 posts
Jun 22, 2017 06:06 PM | Stephen L. - Coma Science Group, GIGA-Consciousness, Hospital & University of Liege
How to denoise with global correlation GCOR
Hello there!
I can't find how to generate GCOR for denoising. Do we need to use a specific preprocessing step to generate it? Or do we simply regress QA_GCOR as a second-level covariate?
Thanks a lot in advance!
I can't find how to generate GCOR for denoising. Do we need to use a specific preprocessing step to generate it? Or do we simply regress QA_GCOR as a second-level covariate?
Thanks a lot in advance!
Jun 28, 2017 12:06 PM | Jeff Browndyke
RE: How to denoise with global correlation
My understanding is that to affect GCOR during preprocessing, you
need to add the GM as a 1st level covariate prior to
denoising. I've been wondering about what the QA_GCOR would
do as a 2nd level covariate, as I would like to control for
potential global signal differences between conditions/timepoints
without actually doing global signal regression at the 1st
level. I don't think they are the same thing.
Jeff
Jeff
Nov 16, 2017 03:11 PM | Stephen L. - Coma Science Group, GIGA-Consciousness, Hospital & University of Liege
RE: How to denoise with global correlation
Thank you Jeff for the input, yes at last CONN workshop I heard
that gCor was something different to global signal regression, and
that we could use it to correct for multi-centric analyses on
different scanners, but I cannot find how?
In addition, I can't seem to be able to generate QA_GCOR in my latest analyses, so I'm not really sure what are the conditions for QA_GCOR to be generated? And is this the covariate we need to regress (but it does not appear in Denoising tab since it's a QA_* variable)?
Any new on this? Does somebody know how to enable gCor regression?
In addition, I can't seem to be able to generate QA_GCOR in my latest analyses, so I'm not really sure what are the conditions for QA_GCOR to be generated? And is this the covariate we need to regress (but it does not appear in Denoising tab since it's a QA_* variable)?
Any new on this? Does somebody know how to enable gCor regression?
Nov 19, 2017 11:11 PM | Stephen L. - Coma Science Group, GIGA-Consciousness, Hospital & University of Liege
RE: How to denoise with global correlation
I think I've found how to do it:
* Do your CONN processing as usual, you need CONN >= 17f.
* At 2nd-level analysis, you should see a QA_GCOR_[session] variable. You can select it and regress it out, just like any other.
* I also added a "multicentric" 2nd-level covariate to model which subject belonged to which center.
I tried to regress both QA_GCOR and multicentric covariates, and I did a sanity check on controls from 2 different centers (with 2 different but close TR: 2.0 and 2.46). There were major differences, like more than between the patients vs controls!
Is there anything else that can be done to adjust for multicentric analysis in CONN? Thanks a lot in advance!
* Do your CONN processing as usual, you need CONN >= 17f.
* At 2nd-level analysis, you should see a QA_GCOR_[session] variable. You can select it and regress it out, just like any other.
* I also added a "multicentric" 2nd-level covariate to model which subject belonged to which center.
I tried to regress both QA_GCOR and multicentric covariates, and I did a sanity check on controls from 2 different centers (with 2 different but close TR: 2.0 and 2.46). There were major differences, like more than between the patients vs controls!
Is there anything else that can be done to adjust for multicentric analysis in CONN? Thanks a lot in advance!
May 8, 2018 01:05 PM | Ines Del Cerro - IDIBELL - Bellvitge Biomedical Research Institute
RE: How to denoise with global correlation
Dear Stephen, Dear experts,
Thank you for sharing your method. I am at the same point as you was.
Did you finally find a solution for this problem? I have been doing some tests using QA_GCOR variable and another 'site' variable (is a categorical with 3 levels; 3 different scanner sites) as regressors in 2nd level.
I get the same results when I compare a model with 'site' regressor, and with 'site' + QA_GCOR variables. But they are a little bit differents when I only introduce QA_GCOR as regressor.
My question is if you think I could be over-correcting using 'site' + QA_GCOR? Do you know a better way to do it?
Thank you in advance!
Best,
Ines
Originally posted by Stephen L.:
Thank you for sharing your method. I am at the same point as you was.
Did you finally find a solution for this problem? I have been doing some tests using QA_GCOR variable and another 'site' variable (is a categorical with 3 levels; 3 different scanner sites) as regressors in 2nd level.
I get the same results when I compare a model with 'site' regressor, and with 'site' + QA_GCOR variables. But they are a little bit differents when I only introduce QA_GCOR as regressor.
My question is if you think I could be over-correcting using 'site' + QA_GCOR? Do you know a better way to do it?
Thank you in advance!
Best,
Ines
Originally posted by Stephen L.:
I think I've found how to do it:
* Do your CONN processing as usual, you need CONN >= 17f.
* At 2nd-level analysis, you should see a QA_GCOR_[session] variable. You can select it and regress it out, just like any other.
* I also added a "multicentric" 2nd-level covariate to model which subject belonged to which center.
I tried to regress both QA_GCOR and multicentric covariates, and I did a sanity check on controls from 2 different centers (with 2 different but close TR: 2.0 and 2.46). There were major differences, like more than between the patients vs controls!
Is there anything else that can be done to adjust for multicentric analysis in CONN? Thanks a lot in advance!
* Do your CONN processing as usual, you need CONN >= 17f.
* At 2nd-level analysis, you should see a QA_GCOR_[session] variable. You can select it and regress it out, just like any other.
* I also added a "multicentric" 2nd-level covariate to model which subject belonged to which center.
I tried to regress both QA_GCOR and multicentric covariates, and I did a sanity check on controls from 2 different centers (with 2 different but close TR: 2.0 and 2.46). There were major differences, like more than between the patients vs controls!
Is there anything else that can be done to adjust for multicentric analysis in CONN? Thanks a lot in advance!
May 9, 2018 12:05 PM | Stephen L. - Coma Science Group, GIGA-Consciousness, Hospital & University of Liege
RE: How to denoise with global correlation
Dear Ines,
Unfortunately, I did not find any other way to regress out the effect of heterogenous scanners settings. For the moment, I dropped this kind of analysis in CONN as I got unreliable and meaningless results doing so, and I am rather doing separate analyses (one per center) that I then combine using a meta-review kind of analysis such as with the PVM software (see Costafreda, S. G. (2009). Pooling fMRI data: meta-analysis, mega-analysis and multi-center studies. Frontiers in neuroinformatics, 3, 33).
I am also looking into other approaches such as ICA regression and machine learning but they are all experimental, the most reliable way currently from what I have read so far seems to be the meta-analysis approach, as it assumes independence between the analyses, it is statistically perfectly fine to compare multicentric analyses this way (just like one can compare multiple studies results).
Unfortunately, I did not find any other way to regress out the effect of heterogenous scanners settings. For the moment, I dropped this kind of analysis in CONN as I got unreliable and meaningless results doing so, and I am rather doing separate analyses (one per center) that I then combine using a meta-review kind of analysis such as with the PVM software (see Costafreda, S. G. (2009). Pooling fMRI data: meta-analysis, mega-analysis and multi-center studies. Frontiers in neuroinformatics, 3, 33).
I am also looking into other approaches such as ICA regression and machine learning but they are all experimental, the most reliable way currently from what I have read so far seems to be the meta-analysis approach, as it assumes independence between the analyses, it is statistically perfectly fine to compare multicentric analyses this way (just like one can compare multiple studies results).
May 10, 2018 02:05 PM | Ines Del Cerro - IDIBELL - Bellvitge Biomedical Research Institute
RE: How to denoise with global correlation
Dear Stephen,
Thank you very much for your kind response. I will take a look at the paper and method you comment.
Best,
Ines
Originally posted by Stephen L.:
Thank you very much for your kind response. I will take a look at the paper and method you comment.
Best,
Ines
Originally posted by Stephen L.:
Dear Ines,
Unfortunately, I did not find any other way to regress out the effect of heterogenous scanners settings. For the moment, I dropped this kind of analysis in CONN as I got unreliable and meaningless results doing so, and I am rather doing separate analyses (one per center) that I then combine using a meta-review kind of analysis such as with the PVM software (see Costafreda, S. G. (2009). Pooling fMRI data: meta-analysis, mega-analysis and multi-center studies. Frontiers in neuroinformatics, 3, 33).
I am also looking into other approaches such as ICA regression and machine learning but they are all experimental, the most reliable way currently from what I have read so far seems to be the meta-analysis approach, as it assumes independence between the analyses, it is statistically perfectly fine to compare multicentric analyses this way (just like one can compare multiple studies results).
Unfortunately, I did not find any other way to regress out the effect of heterogenous scanners settings. For the moment, I dropped this kind of analysis in CONN as I got unreliable and meaningless results doing so, and I am rather doing separate analyses (one per center) that I then combine using a meta-review kind of analysis such as with the PVM software (see Costafreda, S. G. (2009). Pooling fMRI data: meta-analysis, mega-analysis and multi-center studies. Frontiers in neuroinformatics, 3, 33).
I am also looking into other approaches such as ICA regression and machine learning but they are all experimental, the most reliable way currently from what I have read so far seems to be the meta-analysis approach, as it assumes independence between the analyses, it is statistically perfectly fine to compare multicentric analyses this way (just like one can compare multiple studies results).