Paper: Learning Cortical Parcellations Using Graph Neural Networks

[Aly Kotb]

Hello All,

I have two questions on this paper:

1- In this paper https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.797500/full: , the connectivity matrix between each vertex in the cortical surface mesh and each ROI in the destrieux Atlas was computed. I want to make sure that this was done using the -cifti-average-roi-correlation command in the connectome workbench, if not please share with me any codes you have if possible.

2- I do not know also if the parcellation labels per subject can be made available that were used in training the GNNs

Thanks in advance.

[Tim Coalson]

It would probably be better to use -cifti-parcellate and then -cifti-cross-correlation, making the resulting file easier to display in wb_view.  On the other hand, if individual parcellations were used, you might need to alter the parcel mapping in order to average it across subjects afterwards.  I'm not sure why someone would use a gyral-based atlas for averaging fMRI data.

We are working on an improved individual parcellation classifier, and have not publicly released the per-subject outputs of the old method.

Tim

--
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 visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/66fa3ee5-cf04-462e-bba4-21ac9d0aca0an%40humanconnectome.org.

[Glasser, Matthew]

Don’t you mean -cifti-correlation?

Matt.

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/CANm3QU5zPhZRK64iNZ58JWqRzucndJWJG1pWsK5OE5bsYADMpw%40mail.gmail.com.

 

---

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.

[Tim Coalson]

The question and the paper indicated correlation between vertex timeseries and ROI average timeseries, resulting in dpconn or pdconn.  I meant -cifti-cross-correlation, though perhaps I should have specified between the original dtseries and the parcellated ptseries.

For clarity, -cifti-average-roi-correlation should also be able to do this, but its output uses scalars (named maps) instead of parcels (names and lists of vertices/voxels) for the ROIs dimension, and thus requires selecting the ROI by name in wb_view (instead of clicking on the ROI/vertex of interest).  It doesn't allow the use of separate ROIs per input cifti file, so you'd still have to run it per-subject for any individualized parcellation.

Tim

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/C847275B-8E8C-42DE-B2FD-D44BA75D197D%40wustl.edu.

[Aly Kotb]

Cifit-correlation computes the correlation in vertex-vertex manner or ordinate-ordinate resulting in a 91,282 x 91,282 matrix which cannot be practical to train an AI model, instead the paper computed vertex-ROI correlation resulting 64K(number of vertices)-150 (number of ROIs in destrieux),.

[Aly Kotb]

May you please explain further what is meant by altering the parcel mapping in order to average it across subjects afterwards?

Also, as for they gyral atlas they repeated their experiment by using the (Yeo-17) atlas which is functionally aware, but results did not differ, maybe this is due to the ability of the deep learning model to capture the hidden information, as at the end the regions in an ROI defined based on gyral patterns have some degree of functional similarity that was enough for the model to understand in which parcel the vertex should be located.

Aly.

[Aly Kotb]

Something else please, in the CIFTI file are the 64k cortical vertices the first 64K elements as per the order of the array?

[Tim Coalson]

wb_command -cifti-parcellate puts the list of vertices (and/or voxels) each parcel used into the XML part of the header.  But, when using parcellations that are individualized, these lists of vertices change per-subject.  Some wb_command -cifti-* commands, when operating on multiple cifti files, will check for whether the mappings match (if you are doing some ROI analyses, and you have two different subsets with only 3 parcels in each, but they aren't the same areas, this check will stop you from accidentally averaging/subtracting/etc two files that use different areas).

The downside is that wb_command can't figure out when parcels are "conceptually" the same across subjects, despite having lists of vertices that aren't a perfect match, and it will also error in that case.  To fix it, you can overwrite the parcels mapping in the files (say, with the group-consensus parcel definitions) before averaging them.  The -cifti-change-mapping command can make this change.

In the standard "91k grayordinates", the cortical vertices are the first ~59k of the file, not 64k.  This is because we exclude the medial wall vertices.  However, surface-only cifti files are not standardized in this way, some have 59k indices, others have 64k.  It is also possible to generate a surface and volume cifti file that doesn't exclude the medial wall, and has ~96k total indices.  I recommend using wb_command -cifti-separate if you want to get just the cortical data without any complications (there is also -cifti-create-dense-from-template if you want to go straight from cifti to cifti).  You can also use nibabel to get the full indexing information in python, or cifti-matlab in matlab, and extract it that way.

Tim

[Aly Kotb]

Thank you so much for the thorough explanation.

[Aly Kotb]

Another question, when I print the labels table of "L.aparc.a2009s.32k_fs_LR.label.gii'", I do not understand what is the highlighted ???

Also when I compute the correlation between CIFTI files of the sessions the same subject it is around 0.8.  but after I compute the connectivity matrices between vertices and brain regions the correlation between the connectivity matrices of intra subject sessions is around 0.18-0.2!!!

Label table for ALL maps KEY NAME RED GREEN BLUE ALPHA 0 ??? 0.000 0.000 0.000 0.000 1 L_G_and_S_frontomargin 0.090 0.863 0.235 1.000 2 L_G_and_S_occipital_inf 0.090 0.235 0.706 1.000 3 L_G_and_S_paracentral 0.247 0.392 0.235 1.000 4 L_G_and_S_subcentral 0.247 0.078 0.863 1.000

[Glasser, Matthew]

??? is unlabeled.

 

fMRI data have a lot of unstructured noise, particularly single vertices.  A parcels-by-parcels matrix will be more similar.  Also I am not clear which files you are using.

Matt.

 

From: Aly Kotb <alyko...@gmail.com>
Reply-To: "hcp-...@humanconnectome.org" <hcp-...@humanconnectome.org>
Date: Saturday, September 6, 2025 at 6:35 AM
To: HCP-Users <hcp-...@humanconnectome.org>
Cc: Aly Kotb <alyko...@gmail.com>, "tim.c...@gmail.com" <tim.c...@gmail.com>, HCP-Users <hcp-...@humanconnectome.org>
Subject: Re: [hcp-users] Paper: Learning Cortical Parcellations Using Graph Neural Networks

 

 

Another question, when I print the labels table of "L.aparc.a2009s.32k_fs_LR.label.gii'", I do not understand what is the highlighted ???

To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/8c0d8f81-8d40-4735-af82-8cce8422e2d4n%40humanconnectome.org.

[Aly Kotb]

This:   L.aparc.a2009s.32k_fs_LR.label.gii
the suffix of the parcellation result of the left cortex of any subject based on the destrieux atlas generated from FreeSurfer.

[Glasser, Matthew]

The fMRI files.

[Aly Kotb]

1- They are in this directory: HCP_1200/subject id (e.g. 100307)/MNINonLinear/fsaverage_LR32k/   ,

2- Also, this raises a challenge of how you would unify the size of matrices representing the cortical functional time-series of all subjects, as the number of single vertices (noisy) varies from a subject to another, in other words if I am to train a deep learning model am I to include or exclude these vertices, and if to exclude as mentioned how I unify the tensor size??!!

[Aly Kotb]

I've done some search, if I set their connectivity matrix as zero, and cut their edges in a graph representation, I think this would work.