General Departmental Seminar Series
Characterizing and Improving Cluster Size Inference in
Functional Neuroimaging Data Analysis
Thomas Nichols, Ph.D., Assistant Professor,
Department of Biostatistics, University of Michigan
Friday, October 31, 2003, 12-1 p.m.
G5/136 Clinical Science Center, 600 Highland Ave.
Functional brain imaging techniques allow neuroscientists to literally peer into the workings of the human brain. Positron Emission Tomography (PET) and Functional Magnetic Resonance Imaging (fMRI) are two such techniques that allow the imaging of blood flow change, which indicates changes in neuronal activity. Such data are massive, with as many as 1,000 multivariate observations, each with 100,000 elements, and have a complicated spatiotemporal dependence structure.
The usual approach is a massively univariate analysis, where models are fit independently at each point in the brain. The resulting statistic images can be assessed voxel-wise or cluster-wise. Clusters are defined as suprathreshold contiguous voxels. Under the null hypothesis there should be no structure and hence large clusters are evidence of activation.
I will report on three investigations on cluster size inference using nonparametric permutation methods, first two directly compare random field theory (RFT) results to permutation results. First we investigated the validity of random field cluster size inference over a range of smoothnesses, since a smoothness assumption is necessary for a discretely sampled image to behave like a continuous random field. Second we propose a nonparametric version of a cluster size test which is valid under nonstationarity, an assumption which is rarely tenable in practice due to heterogeneous smoothness. Lastly we explore joint inference on cluster-size and statistic intensity within the cluster; we consider several families of combining functions and find none dominate one another, and thus propose a meta-combining statistic which pools information from a collection of joint test statistics. Our results show that, in addition to their well-known validity and flexibility, permutation methods can exhibit improved sensitivity relative to comparable parametric methods.
Return to seminar list