On spatial selectivity and prediction across conditions with fMRI

Functional neuroimaging data are currently routinely used to better understand cognitive processes. They rely heavily on previous findings to formulate hypotheses and narrow the search space to regions of interest (ROIs), most often reported as coordinates of activation peaks [1], or from coordinate databases such as BrainMap [2]. However, understanding the literature is increasingly difficult, so that there is a need for more systematic methods, which use the images themselves to characterize the functional specificity of brain regions [3]. Transfer learning is a method that trains a classifier to learn a discriminant model on a source task, and then generalizes on a target task without further training. It can yield insights on some brain mechanisms if the tasks share specific common effects in some brain regions [4]. The goal of this work is to investigate the power of transfer learning procedures applied to pairs of cognitive contrasts, where the discrimination ability of the classifier quantifies the information shared between brain maps, and thus characterizes at which spatial scale functional contrasts can be jointly classified. We show that in many cases, transfer learning gives poor results in terms of spatial selectivity. To address this limitation, we introduce selection transfer, i.e. classification of brain states on the target task following the canonical procedure [5], but using regions defined on the source task.