Nonlinear component analysis as a kernel eigenvalue problem
Neural Computation
Independent component analysis: algorithms and applications
Neural Networks
Cluster Analysis of Biomedical Image Time-Series
International Journal of Computer Vision
Non-negative Matrix Factorization with Sparseness Constraints
The Journal of Machine Learning Research
Topographic Independent Component Analysis
Neural Computation
A toolbox for model-free analysis of fMRI data
ICA'07 Proceedings of the 7th international conference on Independent component analysis and signal separation
Extended sparse nonnegative matrix factorization
IWANN'05 Proceedings of the 8th international conference on Artificial Neural Networks: computational Intelligence and Bioinspired Systems
Beyond cross-domain learning: Multiple-domain nonnegative matrix factorization
Engineering Applications of Artificial Intelligence
| Hi-index | 0.00 |
In the last decades, functional magnetic resonance imaging (fMRI) has been introduced into clinical practice. As a consequence of this advanced noninvasive medical imaging technique, the analysis and visualization of medical image time-series data poses a new challenge to both research and medical application. But often, the model data for a regression or generalized linear model-based analysis are not available. Hence exploratory data-driven techniques, i.e. blind source separation (BSS) methods are very popular in functional nuclear magnetic resonance imaging (fMRI) data analysis since they are neither based on explicit signal models nor on a priori knowledge of the underlying physiological process. The independent component analysis (ICA) represents a main BSS method which searches for stochastically independent signals from the multivariate observations. In this paper, we introduce a new kernel-based nonlinear ICA method and compare it to standard BSS techniques. This kernel nonlinear ICA (kICA) overcomes the restrictions of linearity of the mixing process usually encountered with ICA. Dimension reduction is an important preprocessing step for this nonlinear technique and is performed in a novel way: a genetic algorithm is designed which determines the optimal number of basis vectors for a reduced-order feature space representation as an optimization problem of the condition number of the resulting basis. For the fMRI data, a comparative quantitative evaluation is performed between kICA with different kernels, nonnegative matrix factorization (NMF) and other BSS algorithms. The comparative results are evaluated by task-related activation maps, associated time courses and ROC study. The comparison is performed on fMRI data from experiments with 10 subjects. The external stimulus was a visual pattern presentation in a block design. The most important obtained results in this paper represent that kICA and sparse NMF (sNMF) are able to identify signal components with high correlation to the fMRI stimulus, and kICA with a Gaussian kernel is comparable to standard ICA algorithms and even more, it yields spatially focused results.