fMRI group analysis with Spatial Bayesian Variable Selection

Introduction: In recent times, Bayesian approaches have been increasingly popular in fMRI data analysis. One obvious appeal of the Bayesian approach is its interpretability. Instead of performing a hypothesis based inference at each voxel with an artificial threshold for declaration of activation, in the Bayesian approach we simply estimate the posterior probability of a voxel being active based on a suitable prior. In fact, with the Bayesian approach the arbitrary thresholding process may actually be eliminated. More importantly, the Bayesian approach permits us to use a framework that can incorporate anatomical information or other expert knowledge into the model with appropriate priors. In a classical framework this can be achieved only with segmentation or a region of interest (ROI) based approach, which is too restrictive. One popular Bayesian approach that does not suffer from the problems of the classical approach is the spatial Bayesian variable selection (SBVS) framework introduced by Smith et al. [1]. However, the method as introduced by Smith et. al. is only applicable to a single subject analysis and hence is of limited interest. In this paper we build upon their model for a single subject and extend it to data from a hierarchical experiment (e.g., group data) which allows inferences to be generalized to a larger population. Furthermore, through this model we can account for anatomical heterogeneity across subjects. To our knowledge, this is the only model that can handle such feature.