Measuring variability of hemodynamics responses in event-related BOLD data using ICA
8th Joint Symposium on Neural Computation, Salk Institute, La Jolla, May 19, 2001 Jeng-Ren Duann, T.P. Jung, W.J. Kuo, T.C. Yeh, S. Makeig, J.C. Hsieh, T.J. Sejnowski.
Computation Neurobiology Lab, Salk Institute Institute for Neural Computation, UCSD IBRU, Taipei-Veterans General Hospital, Taipei, Taiwan Salk Institute CNL 10010 N. Torrey Pines Rd. La Jolla, CA 92037Most current methods for analyzing functional magnetic resonance imaging (fMRI) data assume a priori knowledge of the shape of the hemodynamic response (HR) to experimental stimuli or events in brain areas of interest. In addition, they typically assume homogeneity of both the HR and the non-HR "noise" signals across brain regions and across similar experimental events. Such approaches are inadequate when HRs are unpredictable or vary unpredictably from area to area, or even from trial to trial.
Here we use a data-driven method, an infomax Independent Component Analysis (ICA), to detect and visualize single-trial HRs in event-related fMRI data. ICA assumes that the recorded time course at each brain or non-brain voxel is the sum of a number of activities that are expressed with different strengths (and, possibly, polarities) across different (but possibly overlapping) voxel subsets. Different "component" processes have separable time courses and stable spatial structure (McKeown et al., Human Brain Mapping, 1998). ICA further assumes that the component processes are maximally spatially independent. Each ICA component can be represented by a spatially-fixed three-dimensional component map and an associated activity time course, and represents a source of signal variance affecting a compact or multifocal brain area of linearly dependent influence.
Six subjects participated in four fMRI sessions in which ten bursts of 8-Hz flickering-checkerboard stimulation were presented at 30-s intervals. In different sessions, burst durations were either 0.5 s (short) or 3 s (long). Five axial slices were acquired by a Bruker 3T MR imager with echo-planar imaging (EPI) pulse sequence (TR = 500 ms; TE = 70 ms; flip angle = 90 degrees; matrix = 64 x 64; FOV = 250 x 250 mm; slice thickness = 5 mm with 2 mm gap). ICA decomposition of the resulting blood level oxygenation (BOLD) data revealed independent components with differing stimulus-locked HRs active in primary visual (V1) and medial temporal (MT/V5) cortices respectively.
Contrary to expectation, HRs of the V1 components often exhibited two successive positive peaks in response to short-stimulus bursts, while nearly identical component maps were associated with single-peaked HRs in long-stimulus sessions. HRs exhibited substantial variations across trials, sessions, subjects and brain areas making it impossible for hypothesis-driven methods using fixed model template functions to extract the HR time courses for different brains and brain areas. ICA, combined with single-trial visualization, on the other hand, can reveal dramatic and unforeseen task-related HR features that could easily be otherwise ignored by researchers performing analyses based on an a priori model.