Abstract
A fast exploratory framework for extracting cardiac noise signals contained in rest-case fMRI images is presented. Highly autocorrelated, independent components of the input time series are extracted by applying Canonical Correlation Analysis in the time domain. A close correspondence between some of these components and cardiac noise contributions is established. Our analysis is carried out without using any external monitoring of the subject or any modification applied to the standard image acquisition protocol. Using the results as a priori information about the presence of corrupting cardiac noise, several approaches are suggested that could improve the performance of activation detection algorithms on non-rest-case datasets.
Motivation
During one imaging session, a collection of experiments are run, and this series is preceded and followed by a set of rest-case scans. For these rest-case acquisitions, there are no stimuli presented to the subject and there are no tasks to be executed. While response signals are missing, physiological contributions still manifest in these images. If these are detectable, our goal is to gain information about their nature and spatial origin. That information could be very valuable in subsequent non-rest-case data analysis. Ad hoc experiments suggest the presence of two principal components contributing to the rest-case fMRI signals. Displayed in the frequency domain, a higher (appearing in the .08-.2 Hz frequency range) and a lower, 1/f-type frequency component can be often distinguished. We conjecture that the higher frequency components can be associated with aliased cardiac signals. Whenever they are absent, it means that the aliased cardiac contributions occupy the same lower frequency range as the flicker noise. We verify this hypothesis by using external cardiac measurements and by characterizing the noise sources with respect to the corresponding anatomical dataset.
Lawrence Panych panych at bwh.harvard.edu
William Wells sw at csail.mit.edu
Lilla Zöllei lzollei at csail.mit.edu
Back to the MIT AI Lab page.