1Dept. Bio & Brain Engineering,
KAIST, Daejon, Korea, Republic of
Statistical
parametric mapping (SPM) is widely used for the statistical analysis of brain
activity with fMRI. However, if the general linear model employs a fixed form
of a canonical HRF, the ignorance of experimental and individual variance can
lead to inaccurate detection of the real activation area. A variety of
data-driven methods, which combine independent component analysis (ICA) with statistical
analysis of fMRI dataset, were suggested to overcome the problem, such as the
`HYBICAapproach and the unified `SPM-ICAmethod. However, recent study
demonstrates that representation of the brain fMRI using sparse components is
more promising rather than independent components. Also, the real brain fMRI
signal may be regarded as a combination of small set of dynamic components,
where each of them has different signal patterns and sparsely distributed in
each voxel. Hence, we employ the K-SVD, a powerful sparse dictionary learning
algorithm, to decompose the neural signal into dictionary atoms with specific
local responses. Using the trained sparse dictionary as a design matrix in
SPM, we extract which signal components contribute to the neural activation.
We show the proposed method adapts the individual variation and extract the
activation better than conventional methods.
