Large-scale meta-analysis of functional MRI data

PROJECT SUMMARY The human functional neuroimaging literature has experienced sustained and explosive growth, making it difficult to effectively glean insights from this vast knowledge base. Previously, we introduced Neurosynth, a web-based framework for automated fMRI synthesis that supports meta-analysis at scale across over 12,000 studies. We subsequently extended the ecosystem by developing a powerful platform for expert-driven data annotation, enabling researchers to define and execute custom image-based meta-analyses. Although the Neurosynth framework has become a widely used resource in the neuroimaging community, the laborious challenge of annotating studies at scale remains. We previously addressed this problem by using frequency-based text-mining techniques to extract meaning from the abstracts of articles, but these methods were unable to make fine-grained distinctions between neuroscientific concepts or encode methodological details. Recent breakthroughs in natural language processing and artificial intelligence promise to extract accurate information from unstructured texts with little to no labeled data using Zero Shot Learning. However, the application of state-of-the-art language models to the neuroimaging literature requires a systematic framework to ensure machine comprehension accurately captures known concepts, and robust infrastructure to retrieve information from the literature at scale. The overarching goal of the present project period is to bridge the gap between manual annotation and fully automated meta-analysis, by extending the Neurosynth framework to leverage state-of-the-art natural language processing models to enable precise large-scale automated meta-analysis. In Aim 1, we will develop an extensible framework for information retrieval and validation from neuroimaging articles, with a focus on features relevant to neuroimaging meta-analysis such as study participant demographics (e.g. sample size, age, phenotype), methodological details, and semantic concepts mapped onto expert-defined ontologies. In Aim 2, we will develop and deploy infrastructure for continuous retrieval and processing of the neuroimaging literature, ensuring the timely acquisition of new articles and transforming our meta-analytic study database into a comprehensive knowledge base. In Aim 3, we will develop an interactive researcher-in-the-loop platform meta-analysis, enabling users to use extracted study information to find and identify studies, create gold-standard meta-analyses, and browse hundreds of pre-generated large-scale meta-analyses, enabling discovery-driven science. Meta-analyses will be automatically updated, transforming them from a static to a “living” scientific product that remains relevant and up-to-date. Realizing these objectives will introduce powerful and flexible tools for synthesizing the neuroimaging literature at a large scale and with unprecedented precision and ease. These tools will be freely and publicly available to anyone with an internet connection, enabling rapid and efficient meta-analysis across a broad range of clinical and basic research questions.