Share Email Print
cover

Proceedings Paper

Blind source separation (BSS) for fMRI analysis
Format Member Price Non-Member Price
PDF $14.40 $18.00
cover GOOD NEWS! Your organization subscribes to the SPIE Digital Library. You may be able to download this paper for free. Check Access

Paper Abstract

The major goal of BSS and ICA (Independent Component Analysis) is to recover the source signals from the sensor observations under the following assumptions: (1) the source signals are statistically independent, and (2) the sensor observations are linear mixtures of source signals. Typically, BSS and ICA are based on higher-order statistics, e.g., 'parallel slices' of the fourth-order cumulant tensor (JADE algorithm) or Kurtosis (FastICA algorithm) for non-Gaussian source signals. A few techniques are based on lower-order statistics, e.g., Temporal Decorrelation SEParation (TDSEP) and Extended Spatial Decorrelation (ESD). Our approach reported in this paper is based on second-order statistics only. The spatial prototype patterns (i.e., the image-wise expanded versions of the region images) are considered as source signals and the images (or the sampled images) are considered as sensor observations. The cross-correlations between independent source signals as well as their spatially shifted versions vanish. The outcomes validate the basic requirements of BSS and ICA and lead to a simultaneous diagnolization of two symmetric correlation matrices of the observations. Then a demixing matrix is generated by standard techniques of numerical linear algebra. We have validated this method on simulated images and fMRI images. The result obtained by applying this method to simulated images indicates that the source signals (i.e., the region images) are correctly separated. The result from fMRI analysis by using this method, compared with the results obtained by using the SPM, demonstrates consistency. These results show the theoretical correctness and computational simplicity of this method.

Paper Details

Date Published: 21 May 2001
PDF: 9 pages
Proc. SPIE 4321, Medical Imaging 2001: Physiology and Function from Multidimensional Images, (21 May 2001); doi: 10.1117/12.428151
Show Author Affiliations
Tianhu Lei, Univ. of Pennsylvania (United States)
Jayaram K. Udupa, Univ. of Pennsylvania (United States)


Published in SPIE Proceedings Vol. 4321:
Medical Imaging 2001: Physiology and Function from Multidimensional Images
Chin-Tu Chen; Anne V. Clough, Editor(s)

© SPIE. Terms of Use
Back to Top