help
help > RE: Help with ICA analysis
Jan 27, 2020 10:01 AM | Alfonso Nieto-Castanon - Boston University
RE: Help with ICA analysis
Hi Johana,
ICA maps (spatial components) for a given network represent how this network is expressed in each subject and condition, and that includes a mixture of within- and between- network patterns of connectivity. If you are comparing between groups these spatial patterns for one particular network of interest (either in the "ICA spatial properties" window or re-inserting the non-thresholded component image as an ROI and then performing seed-to-voxel analyses) it is perfectly reasonable to see differences located within the same network (typically we would interpret that as differences in within-network connectivity) and/or outside this network (typically we would interpret that as differences in between-network connectivity).
Regarding the question about including temporal lags in ROI timeseries, sorry that option is not available in CONN. A related option in CONN is the "temporal expansion" option in first-level analysis. You can select a subset of (or all) sources in change the option that reads "no temporal expansion" to "add first-order derivatives". That will include for each seed not only the raw BOLD timeseries but also the first-order temporal derivative of this BOLD timeseries, which allows the analyses to better accommodate small temporal delays in the associations.
Last, regarding the "ICA parcellation ROI file" question, typically we use these parcellations (e.g. instead of the non-thresholded component maps) when we want look at individual networks in finer spatial detail (e.g. study subcomponents), and/or when we want to create a data-driven parcellation of the entire brain or of a particular region of interest, so yes, it is perfectly fine to use these to help you then better explore potential differences between groups in the networks components/subcomponents that you are interested in. Generally speaking, because ICA analyses are performed across all subjects without any information about group membership, it is often perfectly fine to use ICA results in almost any way possible as a preliminary step to help you define precisely those regions/networks/components that seem to better characterize the particular aspects of the connectome that you want to focus on, and, once you have settled on how to best identify those regions/networks/components, then simply run the corresponding confirmatory second-level analyses (e.g. ROI-to-ROI, seed-to-voxel, ICA spatial properties, etc.) to explore between-group differences in connectivity with those areas.
Hope this helps
Alfonso
Originally posted by Johanna Martin:
ICA maps (spatial components) for a given network represent how this network is expressed in each subject and condition, and that includes a mixture of within- and between- network patterns of connectivity. If you are comparing between groups these spatial patterns for one particular network of interest (either in the "ICA spatial properties" window or re-inserting the non-thresholded component image as an ROI and then performing seed-to-voxel analyses) it is perfectly reasonable to see differences located within the same network (typically we would interpret that as differences in within-network connectivity) and/or outside this network (typically we would interpret that as differences in between-network connectivity).
Regarding the question about including temporal lags in ROI timeseries, sorry that option is not available in CONN. A related option in CONN is the "temporal expansion" option in first-level analysis. You can select a subset of (or all) sources in change the option that reads "no temporal expansion" to "add first-order derivatives". That will include for each seed not only the raw BOLD timeseries but also the first-order temporal derivative of this BOLD timeseries, which allows the analyses to better accommodate small temporal delays in the associations.
Last, regarding the "ICA parcellation ROI file" question, typically we use these parcellations (e.g. instead of the non-thresholded component maps) when we want look at individual networks in finer spatial detail (e.g. study subcomponents), and/or when we want to create a data-driven parcellation of the entire brain or of a particular region of interest, so yes, it is perfectly fine to use these to help you then better explore potential differences between groups in the networks components/subcomponents that you are interested in. Generally speaking, because ICA analyses are performed across all subjects without any information about group membership, it is often perfectly fine to use ICA results in almost any way possible as a preliminary step to help you define precisely those regions/networks/components that seem to better characterize the particular aspects of the connectome that you want to focus on, and, once you have settled on how to best identify those regions/networks/components, then simply run the corresponding confirmatory second-level analyses (e.g. ROI-to-ROI, seed-to-voxel, ICA spatial properties, etc.) to explore between-group differences in connectivity with those areas.
Hope this helps
Alfonso
Originally posted by Johanna Martin:
Dear Conn users,
I still don't have any reply, so I insist by reposting my concerns.
I have a few questions of clarification regarding ICA analyses. I would like to specify that I have explored Conn's forum on many occasions, but I have not found a definite answer to my interrogations.
Here's the scenario: I'm looking at resting-state data from two populations (patients and controls), and I am investigating for different functional network connectivity between these two groups.
I ran an ICA network analysis and selected 6 components of interest. But now the challenge of group comparison for these components has appeared.
For what I read on the forum, if I compare groups directly for a component at the second level analysis, I understood that I get the difference between my groups of the connectivity between the Network/Component and every voxel in the brain. Thus, I should only obtain voxels outside of my Component network, right? Which is the same results that I will obtain if I reinject my ICA component in the ROI (as non-thresholded image, by default) a the setup stage. Is that correct? And if yes ... can you help me interpreting this result ?
Thus, if I need to compare the difference between groups inside the component of interest, what analysis should I run?
Also, if I want to directly compare components to other components, I should use ROI-to-ROI analysis. But, is Conn toolbox integrates a temporal lag in its analysis (like described in Arbabshirani, HBP, 2013)? And if not, is there a way of doing it?
Last question: Is it correct to the use non-binary masks created by Conn by using "ICA parcellation ROI file" for the type of analysis we are interested here?
I hope I haven't lost you on the way with all these questions.
I hope that a conn-artist will help me!!!
Thank you in advance,
Johanna
I still don't have any reply, so I insist by reposting my concerns.
I have a few questions of clarification regarding ICA analyses. I would like to specify that I have explored Conn's forum on many occasions, but I have not found a definite answer to my interrogations.
Here's the scenario: I'm looking at resting-state data from two populations (patients and controls), and I am investigating for different functional network connectivity between these two groups.
I ran an ICA network analysis and selected 6 components of interest. But now the challenge of group comparison for these components has appeared.
For what I read on the forum, if I compare groups directly for a component at the second level analysis, I understood that I get the difference between my groups of the connectivity between the Network/Component and every voxel in the brain. Thus, I should only obtain voxels outside of my Component network, right? Which is the same results that I will obtain if I reinject my ICA component in the ROI (as non-thresholded image, by default) a the setup stage. Is that correct? And if yes ... can you help me interpreting this result ?
Thus, if I need to compare the difference between groups inside the component of interest, what analysis should I run?
Also, if I want to directly compare components to other components, I should use ROI-to-ROI analysis. But, is Conn toolbox integrates a temporal lag in its analysis (like described in Arbabshirani, HBP, 2013)? And if not, is there a way of doing it?
Last question: Is it correct to the use non-binary masks created by Conn by using "ICA parcellation ROI file" for the type of analysis we are interested here?
I hope I haven't lost you on the way with all these questions.
I hope that a conn-artist will help me!!!
Thank you in advance,
Johanna
Threaded View
| Title | Author | Date |
|---|---|---|
| Johanna Martin | Aug 9, 2019 | |
| Johanna Martin | Jan 27, 2020 | |
| Alfonso Nieto-Castanon | Jan 27, 2020 | |
| Johanna Martin | Aug 22, 2019 | |