How to estimate within network rest connectivity from ICA on CIFTI data
Leonardo Tozzi
Dear Experts,
I have conducted an ICA analysis on resting state CIFTI data using MELODIC. My dataset consists of several subjects, each with 4 resting state scans.
I have run the ICA using Multi-session temporal concatenation as implemented in MELODIC. Therefore, the output has spatial maps for each component and a timeseries concatenated across all the scans for each component.
Now, I would like to get a measure of within-network connectivity for each component for each individual scan. My plan is to then average these measurements for each subject across their repeated runs and in the end to have a measure of functional connectivity for each of the networks identified by ICA for each subject.
However, I am a bit confused because I don’t know what is commonly used as a measure of within-network connectivity in the ICA framework, since I get one timeseries per component. Between-network connectivity seems more straightforward, since I could just correlate these components. For task data, I know I can do tensor-ICA which will give me one value per subject per component, but from what I understand this approach is not suited for resting state data since it assumes that all subjects are following the same paradigm.
It is possible that the solution could lie in the dual regression framework, but as far as I know, the FSL function for that does not read CIFTIs. Also, in my case, I am not really interested in the individual spatial maps, but rather in a measure of network connectivity I can compare across subjects.
Sorry if this question is a bit basic, but I am a bit new to ICA methods and would greatly appreciate any advice on how to do this analysis.
Thank you!
Leonardo Tozzi, MD, PhD
Director of Computational Neuroscience & Neuroimaging Program
Stanford Center for Precision Mental Health and Wellness
Stanford University
lto...@stanford.edu | (650) 5615738
Connect with us!
Glasser, Matt
Hi Leonardo,
For “dual regression” we have the RSNRegression function in the HCP Pipelines that uses Weighted Multiple Regression (see Glasser et al., 2016 Nature for explanation and comparison with dual regression). I assume this is spatial ICA. Is it group ICA? I actually rarely run ICA in melodic these days (only in ICA+FIX). I prefer to use ICASSO for group spatial and temporal ICA and code for this is available in the temporal ICA pipeline. In fact, if group spatial ICA and dual regression is what you are really after, you could just run the first three modules of the temporal ICA pipeline (MIGP, group sICA, and weighted regression of the group sICA) to generate your dataset.
As for a “within network” measure of connectivity, I assume that means statistical tests of the spatial maps for given components across subjects. That said, there is a thought that temporal ICA or even something like probabilistic functional modes might be better than spatial ICA for this sort of thing.
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/BYAPR02MB5352DC76CBF6DD59076866CCA91A9%40BYAPR02MB5352.namprd02.prod.outlook.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.
Leonardo Tozzi
Dear Matt,
Thank you very much for your reply.
What I was trying to do was indeed spatial ICA and at the group level. My idea was to find some spatial maps across subjects of the major resting state networks, then to somehow derive a measure of “strength” of each network for each subject. By using tensor ICA I managed to do this for task data, since I get three matrices as output, one of which is a subject*component matrix. However, I realize now that this is not feasible with the way the group ICA decomposes the signal for resting state, because I only get 2 matrices.
Of course, the obvious thing to do would be to take a parcellation with network labels, extract the timeseries within each network and correlate them to get the measure of within-network connectivity. I just wanted to try and get at this with ICA, since I am using ICA for the task already.
I will look into ICASSO, temporal ICA and probabilistic functional modes to see if they serve my purpose. Thank you for pointing me to the HCP scripts, I will start from there. Do any of these methods run on other software?
Thank you very much,
Leonardo Tozzi, MD, PhD
Director of Computational Neuroscience & Neuroimaging Program
Stanford Center for Precision Mental Health and Wellness
Stanford University
lto...@stanford.edu | (650) 5615738
Connect with us!
To view this discussion on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/D83CB4D6-5BFB-4BFC-9D91-0011612495DD%40wustl.edu.
Glasser, Matt
Hi Leonardo,
I don’t understand your follow up question. If you have the current version of the HCP Pipelines installed properly they will work.
To view this discussion on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/BYAPR02MB53523A2982300AA6C3AA3A46A91D9%40BYAPR02MB5352.namprd02.prod.outlook.com.
Leonardo Tozzi
To view this discussion on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/2B04B8B2-1FC2-45F9-9E92-F3D570531BE4%40wustl.edu.
Glasser, Matt
Yes this is all implemented and you just need to do as I suggested below regarding the tICA Pipeline modules.
To view this discussion on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/45c3741f-7025-4a6b-9e81-f2fd54f49f75%40email.android.com.