Regarding ICA

[bathula satwik]

I am currently working with rsFMRI data, being treated as univariate time series. I need to reduce the dimensionality of the data, as per my knowledge ICA is being used in FSL.

 1) Does ICA reduce the data from 91282x1200 to n_componentsx1200 or is it just a clustering algorithm?

2)  Is FastICA being used in FSL?

3)  How are the components being defined and how are the grayordinates in that components being determined? 

[Tim Coalson]

ICA is not a clustering algorithm, it starts with PCA (which is an orthogonal decomposition based on the strongest shared variance), and then recombines the PCA components to maximize independence.  In practice, ICA becomes unstable unless you use a smaller number of components than the PCA dimensionality (PCA usually uses the full data dimensionality, so in theory PCA by itself is lossless), so the ICA has to be run as a dimensionality reduction in order to be stable.

When we need to do a dimensionality reduction for analysis purposes, we generally parcellate based on functional area boundaries, instead of using ICA (or folding/gyrus based parcels).  Doing it this way means you have a decent description of what each index in the reduced dimension means neurobiologically, and it minimizes the mixing of signal from different functional areas.

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 on the web visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/df2e4598-b738-424a-992b-f5e6a8a35446n%40humanconnectome.org.