Comparison of canonical correlation analysis and ICA techniques for fMRI

This paper compares independent component analysis (ICA) and canonical correlation analysis (CCA) applied to functional magnetic resonance imaging (fMRI) data. There has been no systematic comparison of these techniques so far. Two variants of the ICA, Infomax and FastICA, are implemented. The CCA method is investigated according to the signal subspace spanned by two hemodynamic response models: differential Gamma and Balloon models. The criterion for the comparison is the area under receiver operating characteristic (ROC) curve for simulated datasets. This criterion is evaluated for different contrast to noise ratios (CNR). Using a real auditory dataset, the paper also compares the aforementioned algorithms in terms of task-related activation maps. The results indicate the superiority of the CCA for CNRs below 0.75; but as the CNR goes beyond this limit, the ICA with Infomax algorithm outperforms other methods. Furthermore, the use of either differential Gamma or Balloon models in the CCA provides nearly the same performance. The paper results can assist the selection of an appropriate algorithm for fMRI data analysis.