|
Abstract:
|
Functional magnetic resonance imaging (fMRI) shows blood-oxygen-level-dependent (BOLD) in the human brain, a proxy for neural activity. An fMRI experiment produces massive amounts of highly complex data; researchers face significant challenges in analyzing their data (Ashby, 2011). A single subject fMRI data contains a time-series data recorded for each voxel (volumetric pixels) and constitutes more than 100,000 voxels in the brain to measure neural activity associated with a variety of brain functions. Although some can be well modeled, the data may contain noise characters including thermal noise, scanner drift, magnetic field inhomogeneity, head motion and human physiology. For fMRI experiment that involves multiple subjects with stimulus induced clinical features, the data entanglement gets even more convoluted. Under these circumstances, the commonly used mass univariate normal model analyses may not be good enough to separate signals from noise and yield reasonably sound results. Multivariate normal model assumption can be effective and more stable alternative method to handle the spatiotemporal fMRI data in two-step estimation procedures.
|
