Towards a spatio-temporal analysis tool for fMRI data:: An application to depth-from-motion processing in humans

Statistical tools for functional neuro-imaging are aimed at investigating the relationship between the experimental paradigm and changes in the cerebral blood Row. They are usually based on univariate statistical techniques [3], however, since cognitive functions result from interactions, a number of new concepts and multivariate tools have recently been developed [5, 7, 12]. A different approach is to perform Input Variable Selection (IVS), a topic that has recently seen a resurgence of interest in the neural network modelling community. It boils down to the selection of a subset of "prototypical" signals that satisfy best a pre-specified criterion. We perform IVS by considering the data set as originating from a multivariate, attractor-based system, and select the prototypes that best approximate the attractor's dynamics [2, 6, 11]. In this way, we are able to consider both spatial and temporal correlations in one pass. As an example, we apply our technique to a fMRI-study concerning "Depth-from-Motion Processing in Humans" and compare our results with those obtained with the popular SPM technique.