Bayesian regression of functional neuroimages

A Bayesian approach is proposed for statistical analysis of fMRI data sets in a two state (“on-off”) activation study. The approach is based on the Relevance Vector Machine (RVM) regression framework. According to this approach the shape of the activations is a superposition of kernel functions, one at each pixel of the image, and a hierarchical Bayesian model is employed which imposes a sparse representation by selecting a number relevant kernel functions. We have implemented an incremental method for constructing the RVM model and, in addition, we have employed a cross-validation criterion to deal with the problem of kernel width selection. The proposed method allows the accurate estimation of the activation locations when correlated noise is present even at low signal-to-noise ratios. We tested this method using an artificial phantom derived from a previous neuroimaging study with promising results compared with previous approaches.