Skip to content

Add Spearman's rho metrics as replacements for the Dice metrics #1352

@tsalo

Description

@tsalo

Summary

The interpretation for the Dice's similarity metrics seems to be: "Do voxels that strongly express component c also show strong TE-dependent (or TE-independent) evidence for that component?" Higher similarity between the component weight map and the TE-dependence F map indicates that the component is more likely to reflect a TE-dependent signal.

The Dice measures are strongly dependent on the thresholds used for each of the maps. Spearman's rho might be a good alternative that works on the unthresholded maps.

Related to #1337, wherein I was seeing that switching from unstandardized parameter estimates to standardized parameter estimates dramatically reduced the Dice values. This was expected, because the unstandardized parameter estimates confound component signal strength and the scale of the raw data, but it also means that the Dice values aren't very useful anymore (i.e., they mostly end up being zeroes).

Metadata

Metadata

Assignees

No one assigned

    Labels

    TE-dependenceissues related to TE dependence metrics and component selectionenhancementissues describing possible enhancements to the project

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions