A1 Refereed original research article in a scientific journal
Brains and phantoms: An ICA study of fMRI
Authors: Ylipaavalniemi J, Mattila S, Tarkiainen A, Vigario P
Publication year: 2006
Journal: Lecture Notes in Computer Science
Journal name in source: INDEPENDENT COMPONENT ANALYSIS AND BLIND SIGNAL SEPARATION, PROCEEDINGS
Journal acronym: LECT NOTES COMPUT SC
Volume: 3889
First page : 503
Last page: 510
Number of pages: 8
ISBN: 3-540-32630-8
ISSN: 0302-9743
Abstract
Biomedical signal processing is arguably the most successful application of independent component analysis (ICA) to real world data. For almost a decade, its use in connection with functional magnetic resonance imaging (fMRI) has allowed for data-driven analysis, partly removing the constraints for stringent experimental setups, which are often required by traditional methods based on the use of temporal references. Recent studies on the consistency of independent components have resulted in a series of tools enabling a more reliable use of ICA. In particular, it is now rather easy to detect algorithmic overfitting and isolate subspaces of related activation. Yet, often the nature of the components may not be determined unambiguously. Focal fMRI signals, seemingly originating from within a subject's brain and showing physiologically plausible temporal behavior, are typically considered relevant. This paper presents a study, which makes use of a standard homogeneous spherical phantom and shows evidence for artifacts caused by the measuring device or environment, with characteristics that could easily be misinterpreted as physiological. Our results suggest that reliable analysis of fMRI data using ICA may be far more difficult than previously thought. At least, artificial behavior revealed by phantom analysis should be considered when conclusions are drawn from real subject measurements.
Biomedical signal processing is arguably the most successful application of independent component analysis (ICA) to real world data. For almost a decade, its use in connection with functional magnetic resonance imaging (fMRI) has allowed for data-driven analysis, partly removing the constraints for stringent experimental setups, which are often required by traditional methods based on the use of temporal references. Recent studies on the consistency of independent components have resulted in a series of tools enabling a more reliable use of ICA. In particular, it is now rather easy to detect algorithmic overfitting and isolate subspaces of related activation. Yet, often the nature of the components may not be determined unambiguously. Focal fMRI signals, seemingly originating from within a subject's brain and showing physiologically plausible temporal behavior, are typically considered relevant. This paper presents a study, which makes use of a standard homogeneous spherical phantom and shows evidence for artifacts caused by the measuring device or environment, with characteristics that could easily be misinterpreted as physiological. Our results suggest that reliable analysis of fMRI data using ICA may be far more difficult than previously thought. At least, artificial behavior revealed by phantom analysis should be considered when conclusions are drawn from real subject measurements.
UTU Research Portal