NeuroSynth: a new platform for large-scale automated synthesis of human functional neuroimaging data

Tal Yarkoni (University of Colorado - Boulder), Russell A. Poldrack (University of Texas - Austin), Thomas E. Nichols (University of Warwick), David C. Van Essen (Washington University School of Medicine), Tor D. Wager (University of Colorado - Boulder)

A central goal of human neuroimaging research is to map relationships between mind and brain, enabling decoding of cognitive states from brain activity (Norman, Polyn, Detre, & Haxby, 2006; Poldrack, Halchenko, & Hanson, 2009). Previous decoding approaches have focused on discriminating between narrow sets of alternative states, overlooking the vast range of cognitive states human beings can experience. Here we introduce a novel framework called NeuroSynth that uses text mining and meta-analysis techniques to automatically produce accurate mappings between brain activity and a large number of broad cognitive states, providing a valuable new resource for a broad range of neuroimaging applications.

The NeuroSynth framework consists of several components. First, an automated parser was used to extract activation coordinates from published neuroimaging studies. The current database contains nearly 145,000 activation foci drawn from over 4,400 studies. Second, all articles were ‘tagged’ with words and phrases that occur at high frequency within the article text, enabling rapid search and filtering. Third, meta-analysis was used to automatically produce whole-brain meta-analysis maps for over 200 distinct psychological concepts, each based on several hundred studies. Fourth, a naïve Bayes classifier was used to classify the probability of specific concepts or states being present given new patterns of activation as input, enabling open-ended decoding and classification of brain activity in both entire studies and individual subjects.

Validation analyses demonstrated the capacity of the NeuroSynth framework to generate meta-analysis maps that closely replicated previous results for many concepts (e.g., pain, emotion, and working memory). Subsequently, we used NeuroSynth to address two long-standing problems in neuroimaging. First, we addressed the well-known problem of reverse inference (Poldrack, 2006) by quantifying the likelihood of specific psychological processes given observed activation patterns. The results identified numerous cases in which reverse inference analyses identified results discrepant with prior findings (e.g., the lack of selectivity of medial frontal function, and the preferential role of posterior rather than anterior insula in pain). Second, we used a naïve Bayes classifier to successfully ‘decode’ which of several psychological states individual subjects were experiencing given only patterns of brain activity, representing an important step towards open-ended classification of cognitive states based solely on prior knowledge.

Collectively, our results validate a powerful new approach to the large-scale synthesis of neuroimaging findings. The tools we have developed enable researchers to address long-standing problems of central importance in neuroimaging, and can be used for a broad range of decoding and classification applications. To further encourage the use and development of synthesis-oriented approaches to neuroimaging data, we are publicly releasing the NeuroSynth data and code on the web ( as an open resource.

Preferred presentation format: Demo
Topic: Neuroimaging

Document Actions