Exploratory parcellation of fMRI data based on finite mixture models and self-annealing expectation maximization

Abstract

We present a new exploratory method for brain parcellation based on a probabilistic model in which anatomical and functional features of fMRI data are used. The goal of this procedure is to segregate the brain into spatially connected and functionally homogeneous components, and to account for variability of the fMRI response in different brain regions. To achieve this goal, the parcellation algorithm relies on the optimization of a compound criterion reflecting both the spatial and functional structures of the individual brains and hence the topology of the dataset. We employ unsupervised learning techniques to classify the fMRI data in an exploratory fashion through a Finite Mixture Model (FMM) for the distribution of the feature vectors of voxels, where the voxels of each parcel follow a normal density. A self-annealing Expectation Maximization (EM) algorithm is used to fit the FMM to the data. To find, the number of parcels, K, we employ Akaike Information Criterion (AIC). The algorithm is tested on synthetic fMRI data as well as real fMRI data. Applying this method to data from a motor experiment, we were able to find homogeneous and connected regions in the motor cortex and the cerebellum that have been previously found using hypothesis-driven methods. Simulation studies have shown that the parcellation results of our method are more accurate than those of the formerly developed methods. © 2010 Springer-Verlag.