An Optimal Independent Component Analysis Approach for Functional Magnetic Resonance Imaging Data

Traditional Independent Component Analysis (ICA) algorithms are based upon the underlying assumption that data implicitly model the probability density functions of the latent sources as highly symmetric. However, when source data violate these assumption, traditional methods might not work well. We propose an Optimal ICA method to model underlying sources, involving two stages procedure. For the first stage, a traditional ICA method is used to obtain initial source estimates, and then, the density of each channel source is calculated with a kernel estimator. At the second stage, it refitts each source by an adaptive nonlinear function. Our simulation data and fMRI experimental results show that the proposed algorithm can separate a wide range of source signal and improve performance on intrinsic skewed data such as the Brain Plasticity during Lexical Associating Learning data.