An Optimal ICA Algorithm Applied to fMRI DATA

Abstract

Conventional independent component analysis (ICA) algorithms are based on the underlying assumption that the probability density functions of the latent sources are highly kurtotic or symmetric. However, when source data violate the symmetric assumption, conventional ICA algorithms might not work well. According to the idea of kernel density estimation of probability density function, an adaptive density model, which incorporates with the adjusted Infomax algorithm, is proposed. A novel optimal ICA method is then obtained. There are two main steps in the presented algorithm. First, an Infomax algorithm is used to obtain initial independent source estimates, and a kernel estimator technique is utilized to calculate source densities. Second, the sources are refitted with a nonlinear function based on their own characteristics, and more precise results can be obtained. Experimental results show that the optimal ICA algorithm, comparing with the other ICA algorithms (e.g. Extended Infomax, FastICA and JADE), can improve separation performance further by incorporating a priori information into ICA analysis of functional magnetic resonance imaging (fMRI) signals. Moreover, it is worth to notice that the Optimal ICA can obtain some special components of fMRI signals that the other ICA algorithms cannot.