Independent Component Analysis Applied to Fmri Data: a Natural Model and Order Selection

We introduce a framework for the application of independent component analysis (ICA) to functional magnetic resonance (fMRI) data. We present a model for the task with two main sections: data generation (synthesis) and data processing (analysis) and give examples of how such a model can be utilized in fMRI analysis. We assume a generative model for the data involving 1) the signal being measured (hemodynamic response to neuronal activation) and 2) the instrument being used to measure this data (the fMRI scanner). Such a structure, as we show, can be useful in designing and evaluating the performance of the data analysis methods. In the analysis section, we incorporate a data reduction stage within the model such that information theoretic criteria can be used to estimate the number of effective brain sources. After describing the components of the model in detail, we give examples of its potential use. We compare principle component analysis to clustering and demonstrate that clustering may be a more natural approach than principle components analysis for data reduction in fMRI analysis and can improve source localization. We demonstrate simulated results and results from an fMRI experiment in which two task related waveforms are present.