Peaks2Image: Reconstructing fMRI maps from stereotactic coordinates to enhance cognitive meta-analysis

To integrate novel findings and reach a consensus on the neural correlates of cognitive functions, human neuroscientists have relied heavily on the literature to formulate, validate, or refine hypotheses. Since thousands of publications have reported the coordinates of regions identified in brain mapping, the community has largely performed meta-analyses based on these coordinates. A known problem with coordinate-based meta-analyses is that they rely on limited information from each publication, so terms appearing in few publications are not reliably mapped. Recently, the use of public images has stimulated meta-analysis, which has become possible due to the creation of large image repositories. NeuroVault is a prominent example of such a repository. These repositories allow researchers to compare images across publications, leading to image-based meta-analyses. However, image-based meta-analysis faces a major obstacle: the scarcity and inconsistency of image annotation and metadata. In this paper, we propose Peaks2Image, a neural network approach that reconstructs continuous spatial representations of brain activity from peak activation tables. Peaks2Image provides dense brain reconstructions that can be combined with extensive annotations from the neuroscience literature. We use these annotations to train a decoder with textual features as labels. This results in a much broader set of decoded terms than current image-based studies. We validated the decoder using 47,000 annotated images and successfully identified 91 out of 108 terms, with a model that was uniquely trained using the neuroscience literature.