help
help > RE: Voxel-to-Voxel analysis
Aug 21, 2015 06:08 PM | Alfonso Nieto-Castanon - Boston University
RE: Voxel-to-Voxel analysis
Hi Annie,
Regarding the connectome-MVPA measure, for each voxel it represents the pattern of connectivity between this voxel and the rest of the brain using just a few abstract components (computed from a principal component decomposition of the between-subjects&conditions variance in these patterns). While the ILC/RCC/ICC/RCS measures each represents a different "aspect" of these patterns (e.g. ILC captures the degree of local connectivity in this pattern, ICC captures the total strength of this pattern, etc.), the MVPA measure is an empirical method to fully represent the shape of these patterns (or at least some of their strongest features). See for example Whitfield-Gabrieli et al. Brain connectomics predict response to treatment in social anxiety disorder. Molecular Psychiatry, 2015, or Beaty et al. Default and exeutive network coupling supports creative idea production. Scientific Reports 2015 for some examples of use.
Regarding the number of components for each measure, ILC and ICC measures are univariate (a single number), while RCC and RCS are spatial-vector measures (three numbers for volume-based analyses, each representing a spatial dimension), and MPVA is also a multivariate measure (as many components as you wish to extract, in your example 5). Typically, when using any of these measures in your second level analyses you would select all of the associated entries in the 'measures' list (e.g. for MPVA analyses you select all 5 of those MVPA_* measures) and leave the default F-contrast in the 'between-measures contrast' field (i.e. eye(N) for N measures).
In general, for your study, I imagine that the main contrast that you would wish to test would be the group-by-condition interaction (i.e. are there any between-group differences in connectivity that are in turn different for the T1 and T2 conditions). To do that in CONN simply select all three groups in the 'subject-effects' list and enter a [-1 1 0; 0 -1 1] between-subjects contrast, and then select both of your two conditions in the 'conditions' list and enter a [-1 1] between-conditions contrast. If you do not have any a priori hypothesis about possible seeds/ROIs that might show this sort of group-by-condition interaction effects in your study, then any of the voxel-to-voxel measures are a good way to start (also you might want to consider graph-theory analyses and/or whole-brain ROI-to-ROI analyses). If you find a significant effect there, then you typically will want to perform post-hoc analyses to characterize those differences (e.g. look at the effects for each group and condition separately to better understand what is driving that group-by-condition interaction). In the case of MVPA analyses typically the post-hoc analyses are performed by using your significant clusters as new seeds/ROIs and then performing seed-to-voxel analyses using those new ROIs as seeds and looking at those maps one single condition/group at a time. In all other cases the post-hoc analyses are performed simply by repeating your original analyses again looking at those effects one single condition/group at a time.
Hope this helps
Alfonso
Originally posted by Annie Möller:
Regarding the connectome-MVPA measure, for each voxel it represents the pattern of connectivity between this voxel and the rest of the brain using just a few abstract components (computed from a principal component decomposition of the between-subjects&conditions variance in these patterns). While the ILC/RCC/ICC/RCS measures each represents a different "aspect" of these patterns (e.g. ILC captures the degree of local connectivity in this pattern, ICC captures the total strength of this pattern, etc.), the MVPA measure is an empirical method to fully represent the shape of these patterns (or at least some of their strongest features). See for example Whitfield-Gabrieli et al. Brain connectomics predict response to treatment in social anxiety disorder. Molecular Psychiatry, 2015, or Beaty et al. Default and exeutive network coupling supports creative idea production. Scientific Reports 2015 for some examples of use.
Regarding the number of components for each measure, ILC and ICC measures are univariate (a single number), while RCC and RCS are spatial-vector measures (three numbers for volume-based analyses, each representing a spatial dimension), and MPVA is also a multivariate measure (as many components as you wish to extract, in your example 5). Typically, when using any of these measures in your second level analyses you would select all of the associated entries in the 'measures' list (e.g. for MPVA analyses you select all 5 of those MVPA_* measures) and leave the default F-contrast in the 'between-measures contrast' field (i.e. eye(N) for N measures).
In general, for your study, I imagine that the main contrast that you would wish to test would be the group-by-condition interaction (i.e. are there any between-group differences in connectivity that are in turn different for the T1 and T2 conditions). To do that in CONN simply select all three groups in the 'subject-effects' list and enter a [-1 1 0; 0 -1 1] between-subjects contrast, and then select both of your two conditions in the 'conditions' list and enter a [-1 1] between-conditions contrast. If you do not have any a priori hypothesis about possible seeds/ROIs that might show this sort of group-by-condition interaction effects in your study, then any of the voxel-to-voxel measures are a good way to start (also you might want to consider graph-theory analyses and/or whole-brain ROI-to-ROI analyses). If you find a significant effect there, then you typically will want to perform post-hoc analyses to characterize those differences (e.g. look at the effects for each group and condition separately to better understand what is driving that group-by-condition interaction). In the case of MVPA analyses typically the post-hoc analyses are performed by using your significant clusters as new seeds/ROIs and then performing seed-to-voxel analyses using those new ROIs as seeds and looking at those maps one single condition/group at a time. In all other cases the post-hoc analyses are performed simply by repeating your original analyses again looking at those effects one single condition/group at a time.
Hope this helps
Alfonso
Originally posted by Annie Möller:
Hi!
I am searching for information about the best way of analysing my Voxel-to-Voxel-data. Maybe there is somone here who can give me advise.
I want to explore the connectivity on the whole-brain-level without using a-priori-chosen ROI's, and also see if there is a difference in connectivity between 3 groups.
I am not quite sure which of the measures "connectome-MVPA", "ILC", "RCC", ICC" and "RCS" is best for my purposes. Also I can find some info on the latter 4, but not really what "connectome-MVPA" measures. What does the "multivariate representation" include?
Because I didn't know which one to choose I included all of them :) and I wonder why each one of them gives 1, 3 or 5 versions in the list of Voxel-to-Voxel Measures in second-level Results... Is that the different components?
The participants in each group have done a rsfMRI-scan before and after an intervention (one sham-group and two different versions of intervention), 25 participants in total.
I have entered the two sessions and created two conditions called T1 and T2 with Onset = 0 and Duration = Inf for Session 1 and Onset = [] and Duration = [] for Session 2 (the opposite for T2). I also removed the condition "rest" that was there as a default.
I have created 3 extra Second-level-covariates (in addition to the "AllSubjects") with the names of my groups and 1:s and 0:s accordingly.
To clearify my question; How do I best explore between-group differences in connectivity when i hypothesise that there are no differences at T1, but hope that there are at T2?
I have been reading some older forum-posts but can't really make sense of it...
Any help is deeply appreciated!
/Annie
I am searching for information about the best way of analysing my Voxel-to-Voxel-data. Maybe there is somone here who can give me advise.
I want to explore the connectivity on the whole-brain-level without using a-priori-chosen ROI's, and also see if there is a difference in connectivity between 3 groups.
I am not quite sure which of the measures "connectome-MVPA", "ILC", "RCC", ICC" and "RCS" is best for my purposes. Also I can find some info on the latter 4, but not really what "connectome-MVPA" measures. What does the "multivariate representation" include?
Because I didn't know which one to choose I included all of them :) and I wonder why each one of them gives 1, 3 or 5 versions in the list of Voxel-to-Voxel Measures in second-level Results... Is that the different components?
The participants in each group have done a rsfMRI-scan before and after an intervention (one sham-group and two different versions of intervention), 25 participants in total.
I have entered the two sessions and created two conditions called T1 and T2 with Onset = 0 and Duration = Inf for Session 1 and Onset = [] and Duration = [] for Session 2 (the opposite for T2). I also removed the condition "rest" that was there as a default.
I have created 3 extra Second-level-covariates (in addition to the "AllSubjects") with the names of my groups and 1:s and 0:s accordingly.
To clearify my question; How do I best explore between-group differences in connectivity when i hypothesise that there are no differences at T1, but hope that there are at T2?
I have been reading some older forum-posts but can't really make sense of it...
Any help is deeply appreciated!
/Annie
Threaded View
| Title | Author | Date |
|---|---|---|
| Annie Möller | Aug 20, 2015 | |
| Alfonso Nieto-Castanon | Aug 21, 2015 | |
| Annie Möller | Aug 24, 2015 | |
| Alfonso Nieto-Castanon | Aug 24, 2015 | |
| Annie Möller | Aug 28, 2015 | |
| Alfonso Nieto-Castanon | Aug 29, 2015 | |
| Annie Möller | Sep 1, 2015 | |