Functional MRI (fMRI) can be used to visualize the brain structures involved in a cognitive process, allowing us to "see the brain think". Conversely, it is much more difficult to identify the cognitive processes involved when an fMRI image is taken in an unknown context.

In fact, this "decoding" requires a statistical analysis of the brain activity involved in a large number of cognitive processes.

Researchers at the NeuroSpin set out to meet this challenge by exploiting all available statistical maps of fMRI images from the largest data repository available, NeuroVault.

• They started by "labeling" the NeuroVault images with concepts from Cognitive Atlas, a knowledge base for cognition, by drawing on the atlas's metadata.
• They "homogenized" tens of thousands of brain images.
• They trained neural networks to predict the labels for the entire set of these images.

 They were able to distinguish more than 50 classes of cognitive processes, despite the heterogeneity of the training data, without any a priori knowledge of the experimental setting.

 These results demonstrate that image-based meta-analyses can be conducted on a large scale and with minimal manual data processing. Moreover, they now enable the identification of cognitive processes associated with specific brain activities.