A mutual information based framework for the analysis of multiple-subject fMRI data

Functional magnetic resonance imaging (fMRI) is a non-invasive method of obtaining images of neural activity in response to a stimulus. The reverse process, decoding fMRI images to infer the underlying stimulus relating to a single subject, is a challenging area of work. The complexity greatly increases when the decoding process has to be sufficiently generic to cover multiple subjects and studies. In order to decode neural images with high degrees of accuracy, the process relies on an extensive database of neural signatures. This paper proposes a framework to generate the elements of the database. We also outline the process involved in decomposing an fMRI dataset into independent subsets corresponding to the neural signatures of the stimuli. Each subset is described in a standard format, based on mutual information derived from Regions of Interest, which in turn are derived by co-registering with a standard atlas.