A method for analyzing the consistency of independent components was presented and its usefulness was tested in a real fMRI study. Although the method was used with FastICA, it should be relatively straightforward to use it with other ICA and BSS algorithms.
The method works by exploiting the variability of ICA algorithms in a bootstrapping and clustering based approach. The method produces important additional information on the estimated independent components, which helps in interpreting the results, and it allows the detection of sources difficult or impossible to find with ICA alone or even with other bootstrapping methods.
The experiments show that the method works well with real fMRI data and, indeed, makes interpreting the results easier and more reliable. The method verifies the consistency of the expected results, but also reveals unexpected components in a reliable way. Additional information on the variability of the less consistent components helps interpreting the underlying phenomena.
Additionally, the visualization tools are able to give fast and clear overviews of the analysis results, but also allow the indepth study of the most interesting features. This makes interpreting the results even easier.
Consistent analysis of fMRI studies is crucial in the medical field and the presented method offers a way to do exactly that. The additional information on the variability, especially the location, is also interesting. Possibly helping to improve other methods and form new hypotheses on brain function as well.
The added reliability of the interpretations made from results produced with the method is very important from a medical point of view. The proposed method is also considerably faster that other similar methods and makes it more usable in real medical research.
The presented method and tools will be used extensively on future research projects. This could allow further improvements to be made on the clustering and even on running ICA multiple times, when additional experience is gathered. The visualization tools could be made more compatible with other medical tools allowing better integration into existing analysis setups.
The additional information on the source and nature of variability could also be used to improve the original solution (c.f., Friman et al., 2004). The framework of denoising source separation (DSS) (Särelä and Valpola, 2005) allows the use of such information in a highly integrated way during the estimation. Additionally, in DSS the estimation does not have to be based strictly on statistical independence.
When variability with different natures can be identified and analyzed, there is no reason why the method presented could not be extended also for group studies, where a similar problem exists in comparing results across many subjects (c.f., McNamee and Lazar, 2004, Calhoun et al., 2001a,b). Solutions to the problems in group studies are important, since group studies are standard in the medical field.