Independent component analysis (ICA) of functional magnetic resonance imaging (fMRI) data can be employed as an exploratory method. The lack in the ICA model of strong a priori assumptions about the signal or about the noise leads to difficult interpretations of the results. Moreover, the statistical independence of the components is only approximated. Residual dependencies among the components can reveal informative structure in the data. A major problem is related to model order selection, that is, the number of components to be extracted. Specifically, overestimation may lead to component splitting. In this work, a method based on hierarchical clustering of ICA applied to fMRI datasets is investigated. The clustering algorithm uses a metric based on the mutual information between the ICs. To estimate the similarity measure, a histogram-based technique and one based on kernel density estimation are tested on simulated datasets. Simulations results indicate that the method could be used to cluster components related to the same task and resulting from a splitting process occurring at different model orders. Different performances of the similarity measures were found and discussed. Preliminary results on real data are reported and show that the method can group task related and transiently task related components.