NeuroConText: Contrastive Learning for Neuroscience Meta-Analysis with Rich Text Representation
Coordinate-based meta-analysis (CBMA) is a common approach to synthesize information about human brain function across existing literature, enabling researchers to formulate hypotheses and contextualize new findings. However, automated CBMA tools face challenges such as inconsistent terminology, limited ability to analyze long texts, and difficulty capturing semantic meaning, as they still rely on bag-of-words approaches. In addition, sparse coordinate reporting distorts the activation distribution due to incomplete data. This paper introduces NeuroConText, an automated CBMA tool that bridges neuroscience text, brain location coordinates, and brain images by creating a shared latent space for encoding text and brain maps. Our method relies on a multi-objective loss combining contrastive and reconstruction terms. It leverages large language models (LLMs) to extract neuroscientific information from full-text articles and employs an LLM-based text augmentation strategy to improve generalization to short-text inputs. Quantitative and qualitative analyses demonstrate NeuroConText's ability to enhance text-to-brain retrieval performance and reconstruct brain maps from neuroscience texts. We also show that CBMA tools can infer brain activations in regions discussed in articles but absent from reported coordinates, potentially addressing the challenge of sparse coordinate reporting.