varoquaux@normalesuporg Multivariate random effects: group model for ICA analysis of multi-subject fMRI
Multivariate data analysis procedures, such as ICA (independent component analysis) are widely used to extract spatial patterns displaying salient and meaningful features from fMRI datasets. However, the extract maps differ widely when estimated on different groups of healthy controls.
We introduce a multivariate group model to model subject-to-subjet variability and extract reproducible inter-subject ICA patterns.
CanICA = CCA + ICA: a multivariate random effects
We model 2 levels of variance for noise in the fMRI dataset:
- Observation noise confonding subject-specific patterns that we estimate via PCA (principal component analysis)
- Subject-to-subject variability, present in the variability of the group level, as they are expressed in subject-specific patterns. We estimate group-level patterns from subject-specific patterns using CCA (canonical component analysis)
- Independent sources. From the group-level canonical components, we apply blind source separation using ICA to estimate independent sources.
Diagram of the estimation procedure
Reproducibility results on resting-state fMRI
We draw sub-groups from a group of healthy controls and estimate patterns on the sub-groups. We look at the reproducibility of the patterns extracted.
Method One-to-one reproducibility Melodic TensorICA .35 GIFT (Group ICA) .10 Melodic Concat .50 Raw ICA .21 CanICA .62
Please read the publication for details on these scores.
Software
Python code implementing this model is available on: https://github.com/GaelVaroquaux/canica
Best viewed with firefox, or any browser that respects standards.