Gordon333 parcellation, low correlation values and -cifti-label-export-table command

317 views
Skip to first unread message

tali we

unread,
Apr 21, 2021, 5:52:59 AM4/21/21
to HCP-Users
Dear all,

I tried to parcel according to
Gordon333_FreesurferSubcortical.32k_fs_LR.dlabel.nii
for:
rfMRI_REST1_LR_Atlas_MSMAll_hp2000_clean.dtseries.nii

The Pearson correlation within networks and even between hemispheric homologous regions (Right/Left amygdala, hippocampus) are really low! p<0.1

The second question:
I used wb_command -file-information on the parcellated file. 
I like to use -cifti-label-export-table on the dlabel file, I don't understand this "map" input:
 <map> - the number or name of the label map to use

Thank you for your help
Tali

Glasser, Matthew

unread,
Apr 21, 2021, 5:59:10 AM4/21/21
to hcp-...@humanconnectome.org
  1. Did you convert to Z scores by accounting for temporal degrees of freedom (e.g. as done by FSLNets)?  High spatial and temporal resolution fMRI data has a lot of unstructured (i.e. random) noise in it so raw r values can be relatively low and yet correlations are highly significant given the large numbers of timepoints. 
  2. This is probably ‘1’

 

Matt. 

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/9a9a7102-cb5b-4f98-95c1-6c22453d8d8an%40humanconnectome.org.

 


The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.

Glasser, Matthew

unread,
Apr 21, 2021, 4:50:25 PM4/21/21
to Shreyas Harita, HCP-Users

That does not account for the large number of temporal degrees of freedom, so that is a small z, not a big Z. 

 

Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Wednesday, April 21, 2021 at 3:49 PM
To: HCP-Users <hcp-...@humanconnectome.org>
Cc: "Glasser, Matthew" <glas...@wustl.edu>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

can you convert to Z-scores using wb_command -cifti-correlation  --> [-fisher-z] - apply fisher small z transform (ie, artanh) to correlation

from here: https://www.humanconnectome.org/software/workbench-command/-cifti-correlation

tali we

unread,
Apr 23, 2021, 12:38:53 AM4/23/21
to HCP-Users, glas...@wustl.edu, Shreyas Harita
Dear Matt, 

I'm using only one resting state scan (15 min) from each subject, I compare between the first scan from each session.
1. I assume I should ZSCORE the time-course of each parcel before the Pearson and Fisher transformation? (I'm familiar with graph-analysis that applied ZSCORE after)
2. What is the best way to account for autocorrelation in resting state of HCP? (I saw that FSLnets toolbox uses a Monte Carlo approach to estimate the variance of sample correlation coefficients), however I don't know how to applied it myself, Is there a relevant command in wb_command ?

Tali

tali we

unread,
Apr 23, 2021, 1:00:20 AM4/23/21
to HCP-Users, tali we, glas...@wustl.edu, Shreyas Harita
Furthermore, I'm a little bit confused...
In the article: "Investigations into within- and between-subject resting-state amplitude variations" by Janine Bijsterbosch (Neuroimage, 2017), they estimate the "amplitude": the temporal standard deviation across the run.
It feel wrong to  Zscore the time course before calculating the correlation, as the amplitude also consist real neural signal...(and not only random Noise...)

Tali

Glasser, Matthew

unread,
Apr 23, 2021, 6:09:42 AM4/23/21
to tali we, HCP-Users, Shreyas Harita

It is not recommended to use only a single run.  At a minimum one should use balanced amounts of LR and RL data; however, individua subject values will be more stable if you use all 4 runs.

 

We typically variance normalize the unstructured noise across runs and then return the data to its intensity bias corrected form; though whether data is variance normalized or bias corrected overall will not affect correlation values.  If you are interested in amplitudes, this would be done in intensity bias corrected data (which may currently be hard to get in HCP-YA due to the use of the legacy bias correction—we will clean this up in the future when we provide temporal ICA cleaned data).

 

We use FSLNets to get proper big Z stats for correlations that account for both the number of timepoints and autocorrelation.  We don’t “reinvent the wheel” by reimplementing something that has already been implemented nicely in FSLNets.

Glasser, Matthew

unread,
Apr 23, 2021, 10:17:43 AM4/23/21
to Shreyas Harita, HCP-Users

Why not use FSLNets on both?

Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 9:15 AM
To: HCP-Users <hcp-...@humanconnectome.org>
Cc: "Glasser, Matthew" <glas...@wustl.edu>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

 

Thanks for the clarification Matt.

We’ve used the average of 4 resting fMRI runs (RL + LR).

And our correlation values are at ~0.1-0.2.

 

How the patterns/regions of functional connectivity and similar compared to using the 1000 subjects averaged fMRI data ...

 

Is this acceptable to be used in our results so long as we clearly explain why there’s a difference (ie high corr vs low corr) between the two datasets?

 

Thanks,

Shrey

Glasser, Matthew

unread,
Apr 23, 2021, 10:48:55 AM4/23/21
to Shreyas Harita, HCP-Users

I would do both.  I don’t know if anyone has made something like FSLNets in python, but you could ask on the FSL list.

 

Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 9:31 AM
To: "Glasser, Matthew" <glas...@wustl.edu>
Cc: HCP-Users <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

 

Also why would we apply fslnets to the 1000 subj avg fMRI data?

 

 

On Fri, Apr 23, 2021 at 10:22 AM Shreyas Harita <harita....@gmail.com> wrote:

We’re doing our analysis in Python. 

FSLNets is a matlab toolbox.

 

We’re interested in patterns of connectivity rather than the strength (not that connectivity strength is less important in any way) ...

 

Is there a way to import fslnets from matlab to Python?

Glasser, Matthew

unread,
Apr 23, 2021, 11:00:56 AM4/23/21
to Shreyas Harita, hcp-...@humanconnectome.org

If you are referring to the group MIGP results, those are also going to be decreased a little by unstructured noise but not as much as individuals.  It is true that FSLNets modelling will not work with group average MIGP data.   I’m not sure you can convert those easily to big Z values.

 

Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 9:52 AM
To: "Glasser, Matthew" <glas...@wustl.edu>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

 

Sure thing.

And the hcp 1000 subj avg fMRI data gives normal correlation values.

It’s not as low as the subject specific data (rfMRI_REST1_LR_Atlas_MSMAll_hp2000_clean.dtseries.nii)

Thanks Matt.

tali we

unread,
Apr 23, 2021, 11:34:17 AM4/23/21
to HCP-Users, glas...@wustl.edu, Shreyas Harita, tali we
Thank you Matt,

I understand why to ZSCORE for concatenating between runs (I don't really understand how ZSCORE account for autocorrelation).
What can I apply on a single scan (15 min) to account for the autocorrelation?

Anyhow, we use sub-set of subjects (N=100) and only one run as we are looking for an individual physiological data features.
Of course ZSCORE the time course in this case won't change the correlations...
The Pearson correlation are really low for Gordon333 (<0.05, even between left and right amygdala the highest value is 0.07, mean=0.008), much better for Cole (DMN: 0.2-0.3 primary visual: 0.4-0.5...).

Tali

Harms, Michael

unread,
Apr 23, 2021, 11:36:18 AM4/23/21
to hcp-...@humanconnectome.org, Shreyas Harita



What precisely is the problem with the “low” correlation values?  If you are simply going to use then to compare across subjects, or in predictive modeling, then I doubt that converting to “big Z” is really going to change anything.

 

Cheers,

-MH

 

-- 

Michael Harms, Ph.D.

-----------------------------------------------------------

Associate Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.                        Tel: 314-747-6173

St. Louis, MO  63110                          Email: mha...@wustl.edu

 

From: "Glasser, Matthew" <glas...@wustl.edu>
Reply-To: "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Date: Friday, April 23, 2021 at 10:00 AM
To: Shreyas Harita <harita....@gmail.com>, "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

* External Email - Caution *

Glasser, Matthew

unread,
Apr 23, 2021, 11:58:21 AM4/23/21
to tali we, HCP-Users, Shreyas Harita

We variance normalize unstructured noise between runs so that the unstructured noise level is the same across runs.  Then we unvariance normalize across all runs and convert them to intensity bias corrected status so that things like betas or amplitudes mean the same thing across the whole brain.  These things have nothing to do with temporal autocorrelation, that is what FSLNets models.

 

Converting from r to big Z (using the Fischer transform plus the FSLNets modeling) will help with the low correlation values, which are biased low by unstructured noise relative to noise free data. 

Glasser, Matthew

unread,
Apr 23, 2021, 12:08:17 PM4/23/21
to Shreyas Harita, hcp-...@humanconnectome.org

That won’t be biased to lower little z values by unstructured noise, but isn’t Big Z either.

 

Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 11:01 AM
To: "Glasser, Matthew" <glas...@wustl.edu>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

 

1003 subject group avg fMRI data = /HCP_1200/HCP_S1200_1003_rfMRI_MSMAll_groupPCA_d4500ROW_zcorr.dconn.nii

 

This is what I'm referring to. 

 

 

Glasser, Matthew

unread,
Apr 23, 2021, 12:15:41 PM4/23/21
to Shreyas Harita, hcp-...@humanconnectome.org

That sound fine.


Matt.

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 11:13 AM
To: "Glasser, Matthew" <glas...@wustl.edu>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

 

Thank Matt. 

After doing a bit of reading/questioning people more knowledgeable about this than myself, I've decided to keep the low correlation values. 

The reason for this is - I am using the rs-fMRI data to spatial connectivity patterns across subjects and in predicting what the outcome of rTMS stimulation could be. With the scope of the study in mind, I doubt that converting to “big Z” is really going to change anything.

 

Regards,

Shreyas

 

Harms, Michael

unread,
Apr 23, 2021, 12:26:50 PM4/23/21
to Shreyas Harita, HCP-Users

 

It seems like you’ve come to a resolution.  Of course, it is not surprising that the r values from the group dense connectome would be substantially higher than in individual subjects.

 

-- 

Michael Harms, Ph.D.

-----------------------------------------------------------

Associate Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.                        Tel: 314-747-6173

St. Louis, MO  63110                          Email: mha...@wustl.edu

 

From: Shreyas Harita <harita....@gmail.com>
Date: Friday, April 23, 2021 at 11:21 AM
To: HCP-Users <hcp-...@humanconnectome.org>
Cc: "Harms, Michael" <mha...@wustl.edu>, Shreyas Harita <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

* External Email - Caution *

Hi Michael, 

I'll try to keep it brief -

I'm a grad student and my project is looking at inter-individual variability in rTMS target connectivity (e.g. dlPFC). 

 

There are 2 types of rs-fMRI datasets I'm working with -

 

The first is /HCP_1200/HCP_S1200_1003_rfMRI_MSMAll_groupPCA_d4500ROW_zcorr.dconn.nii where the correlation values. This is the average rs-fMRI data across 1003 subjects.

The correlation values here are between ~0.5\0.6

 

The second is rfMRI_REST1/REST2_LR/RL_Atlas_hp2000_clean_cifti_correlated.dconn.nii. These are four resting-state runs (rest 1 || RL/LR phase encoding and rest 2 || RL/LR) for each individual subject. I've averaged the 4 resting-state runs for my project.

The correlation values here are very low at ~ 0.05, max 0.2.

 

However, the patterns of connectivity, are the same between both datasets. Patterns of connectivity = same regions are correlated within the DMN for example, or any other network you wish to look at. 

My only concern is that this discrepancy between the two datasets sets could invalidate my results as I am using two datasets with different correlation values.

The question I am interested in is the difference in the spatial connectivity patterns between a group average fMRI and individual fMRI i.e., do the regions correlated vary. and If they do, how similar/different are they?

 

Thanks for your help.

Regards,

Shreyas

Coalson, Timothy Scott (S&T-Student)

unread,
Apr 23, 2021, 4:54:40 PM4/23/21
to Shreyas Harita, HCP-Users
When you say you averaged across the 4 runs, you don't mean you averaged the timeseries before correlation, right?  Probably not, or the correlation values would be even lower, but just in case...the efficient way to do this is to demean and concatenate the timeseries before correlating.

Tim


From: Harms, Michael <mha...@wustl.edu>
Sent: Friday, April 23, 2021 11:26 AM
To: Shreyas Harita <harita....@gmail.com>; HCP-Users <hcp-...@humanconnectome.org>

Shreyas Harita

unread,
Apr 23, 2021, 5:26:32 PM4/23/21
to HCP-Users, glas...@wustl.edu
can you convert to Z-scores using wb_command -cifti-correlation  --> [-fisher-z] - apply fisher small z transform (ie, artanh) to correlation

from here: https://www.humanconnectome.org/software/workbench-command/-cifti-correlation

On Wednesday, April 21, 2021 at 5:59:10 AM UTC-4 glas...@wustl.edu wrote:

Shreyas Harita

unread,
Apr 23, 2021, 5:26:42 PM4/23/21
to HCP-Users, glas...@wustl.edu, Shreyas Harita
Is there a way to do this without FSLNets? ie are there any alternatives?

Shreyas Harita

unread,
Apr 23, 2021, 5:26:58 PM4/23/21
to HCP-Users, Glasser, Matthew
Thanks for the clarification Matt.
We’ve used the average of 4 resting fMRI runs (RL + LR).
And our correlation values are at ~0.1-0.2.

How the patterns/regions of functional connectivity and similar compared to using the 1000 subjects averaged fMRI data ...

Is this acceptable to be used in our results so long as we clearly explain why there’s a difference (ie high corr vs low corr) between the two datasets?

Thanks,
Shrey
On Fri, Apr 23, 2021 at 6:09 AM Glasser, Matthew <glas...@wustl.edu> wrote:

Shreyas Harita

unread,
Apr 23, 2021, 5:27:09 PM4/23/21
to Glasser, Matthew, HCP-Users
We’re doing our analysis in Python. 
FSLNets is a matlab toolbox.

We’re interested in patterns of connectivity rather than the strength (not that connectivity strength is less important in any way) ...

Is there a way to import fslnets from matlab to Python?

Shreyas Harita

unread,
Apr 23, 2021, 11:14:57 PM4/23/21
to Coalson, Timothy Scott (S&T-Student), HCP-Users
Hey Tim,

No, I didn't average the time series before correlation. Yeah, this bit was a bit tricky, so I decided to average the functional correlation of each run. So parcel X correlated with run 1, 2, 3, 4 and average correlation of this.

Regards,
Shrey

Shreyas Harita

unread,
Apr 23, 2021, 11:15:04 PM4/23/21
to Coalson, Timothy Scott (S&T-Student), HCP-Users
To clarify, I used wb_command -cifti-correlation on the dtseries.nii to get a dconn.nii connectivity matrix. Repeated this for each of the 4 resting state runs and did the above ie., parcel X to matrix 1-4 and mean of this.
Regards,
Shrey

Shreyas Harita

unread,
Apr 23, 2021, 11:15:12 PM4/23/21
to Coalson, Timothy Scott (S&T-Student), HCP-Users
As I've mentioned above, I'm not particularly interested in the actual correlation values. I just needed to make sure that low values are ok for what I'm doing. I'm using the rs-fMRI data to look at spatial connectivity patterns across subjects and in predicting what the outcome of rTMS stimulation could be. 

Apologies for the multiple spams.

- Shrey

Harms, Michael

unread,
Apr 23, 2021, 11:26:11 PM4/23/21
to hcp-...@humanconnectome.org, Coalson, Timothy Scott (S&T-Student)

 

For what you are doing, generating a dense connectome (dconn.nii) seems terribly inefficient.

 

I’d run -cifti-parcellate on the dtseries.nii to generate a ptseries.nii.  And on that run -cifti-correlation to generate the pconn.nii.

 

Cheers,

-MH

 

-- 

Michael Harms, Ph.D.

-----------------------------------------------------------

Associate Professor of Psychiatry

Washington University School of Medicine

Department of Psychiatry, Box 8134

660 South Euclid Ave.                        Tel: 314-747-6173

St. Louis, MO  63110                          Email: mha...@wustl.edu

 

From: Shreyas Harita <harita....@gmail.com>
Reply-To: "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Date: Friday, April 23, 2021 at 10:15 PM
To: "Coalson, Timothy Scott (S&T-Student)" <tsc...@mst.edu>
Cc: HCP-Users <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] Gordon333 parcellation, low correlation values and -cifti-label-export-table command

 

* External Email - Caution *

Glasser, Matthew

unread,
Apr 24, 2021, 8:55:46 AM4/24/21
to hcp-...@humanconnectome.org, Coalson, Timothy Scott (S&T-Student)

That is another reason your correlation values are low.  You are not taking advantage of the SNR improvements from a parcellated analysis.  With fMRI, you  always want to parcellate first and then analyze so that you average out unstructured noise when averaging across parcels.

 

Matt.

Shreyas Harita

unread,
Jun 3, 2021, 10:59:01 AM6/3/21
to HCP-Users, glas...@wustl.edu
Hi Matt, 

You mention that the correlations are 'highly significant', despite the raw r values being low. Do you have an effect size for this? What is the definition of significant here?

Thanks.
Regards,
Shreyas

On Wednesday, April 21, 2021 at 5:59:10 AM UTC-4 glas...@wustl.edu wrote:

Glasser, Matthew

unread,
Jun 3, 2021, 11:01:29 AM6/3/21
to Shreyas Harita, HCP-Users

If you want to compute an effect size you would need to compensate for the biasing effect of random noise.  The ideal case (with a perfect MRI scanner) would be noiseless and so correlations would only deviate from 1 based on the presence of other neural signals.


Matt.

 

From: Shreyas Harita <harita....@gmail.com>


Date: Thursday, June 3, 2021 at 9:48 AM
To: HCP-Users <hcp-...@humanconnectome.org>
Cc: "Glasser, Matthew" <glas...@wustl.edu>

Reply all
Reply to author
Forward
0 new messages