Print

Print
Hi Leonardo,

Please see below:


On Fri, 29 Mar 2019 at 18:37, Leonardo Tozzi <[log in to unmask]> wrote:

Dear Experts,

 

I have a large number of subjects that have undergone the exact same fMRI task. I would like to run an ICA using MELODIC to identify components associated with the task design across all subjects. In particular, I would be interested in the variation of the components in the subject domain and, from what I understand, tensor ICA is suited for this purpose, since it returns a subject*component matrix.

I have two questions. First of all, how can one intuitively conceptualize the elements of this matrix? Is this a weighting factor that represents the individual variation in this component at the subject level?


Yes.

 

Could I then, for example, choose a component that is associated with my task design and correlate the corresponding values of this matrix (one per subject) to task performance or a questionnaire?


Yes. However, you should not compare component strengths in absolute values. That is, like with the usual 2D-ICA, tri-dimensional decomposition cannot preserve the relative variances across components, nor their sign/direction. As long as you look into correlations or t-statistics, and do not make inferences on the direction of the effects, results should be valid and interpretable.

If you look into more than one component, remember to correct for multiple testing, and note that independence is maximal on the spatial domain, but not on temporal or subject domains, such that Bonferroni will be conservative.

 

The second questions pertains to dual regression. From what I understand, it is common practice to use this procedure to assess the differences in spatial extent of components between, for example, two groups. However, if I focused on the individual score mentioned above, would it be necessary to run a dual regression at all? From my understanding, running dual regression would “just” tell me where in the brain the association between the group component and the individual’s timeseries is highest, but I am not sure how the decomposition in 3 matrices of TICA (instead of 2) factors in this procedure (or if it is at all possible). I would imagine that the dual regression would allow me to identify the voxels that show the best fit to the component, while accounting at the same time for the subject score. Is this correct?


You can still do the dual regression. It will give something that tensor ICA doesn't yield: subject-level spatial maps. These can then be subjected to statistical analysis, allowing for localizing power. Interpretation is the same as in the usual dual-regression, except noting that differently than in the temporal concatenation ICA, in which a set of components that have spatial features that are common across participants is produced, here a set of components that have temporal AND spatial features that are common across participant is produced.

Also, I don't think the dual_regression script will run out of the box. You'd need to tweak it yourself...

All the best,

Anderson


 

Thank you very much for your help,

 

 

Leonardo Tozzi, MD, PhD

Williams PanLab | Postdoctoral Fellow

Stanford University | 401 Quarry Rd

[log in to unmask] | (650) 5615738

 


To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1


To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1