ica

Provide a list of preprocessed resting-state BOLD images, an optional group mask, and the desired number of components. The ICA fitting is executed through a nilearn-based reference script that outputs component maps and subject time series.

ICA expects preprocessed resting-state fMRI including nuisance regression, filtering, and standard-space alignment. These steps should be completed first with a dedicated fMRI preprocessing skill before decomposition.

The number of components strongly influences decomposition granularity, with values like 20 being common for network-level analysis. Lower counts yield coarse networks while higher counts split networks into finer subcomponents.

ICA is unsupervised and does not provide statistical group inference by itself. Group comparisons require additional statistical analysis on the extracted subject-level time series or component loadings after decomposition.

Use ICA when the goal is interpretable resting-state network decomposition rather than supervised phenotype prediction. It is a lightweight classical method suited for extracting spatial networks and time series without labeled training data.