Independent component analysis in the presence of noise in fMRI

Dietmar Cordes, Rajesh Nandy
2007 Magnetic Resonance Imaging  
Introduction ICA is a statistical method for estimating a collection of unobservable signals from observations of their mixtures 1 . An explicit treatment of the noise in fMRI data using ICA is difficult primarily due to the complex structure of the noise. For example, noise studies with phantom data have shown that the Fourier spectrum has low frequency components with approximate 1/f dependence suggesting a low-order autoregressive structure. For better source estimation, it may be necessary
more » ... o consider additive colored noise in the ICA model. Due to the complexity in dealing with noise, the application of a noisy ICA model has been neglected. We applied the concept of Gaussian Moments as introduced by Hyvärinen 2,3 to remove the noise-induced asymptotic bias and computed a more accurate mixing matrix in the first stage of ICA. Furthermore, using parametric estimates of the sources densities, we used maximum likelihood estimation to determine the de-noised sources in a second ICA stage.
doi:10.1016/j.mri.2007.03.021 pmid:17509787 fatcat:6zkw6uldz5dvti4qa7pe6xp6hi