NeuroConText: Contrastive learning for neuroscience meta-analysis with rich text representation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Traditional brain meta-analysis struggles with inconsistent terminology and sparse data.
• Existing tools often rely on bag-of-words models, limiting semantic understanding and text analysis.
• Incomplete coordinate reporting in studies distorts brain activation mapping.

Purpose of the Study:

• Introduce NeuroConText, a novel framework for predictive brain meta-analysis.
• Bridge the gap between neuroscience text, brain coordinates, and brain images.
• Enhance the accuracy and comprehensiveness of meta-analytic findings.

Main Methods:

• Developed a predictive text-to-brain modeling framework using a shared latent space.
• Employed contrastive learning for study retrieval and a multi-objective loss for combined retrieval and reconstruction.
• Utilized large language models (LLMs) for full-text analysis and text augmentation for short inputs.
• Integrated NeuroConText with coordinate-based meta-analysis (CBMA) for second-level statistical synthesis.

Main Results:

• Demonstrated enhanced text-to-brain retrieval performance.
• Successfully reconstructed brain maps from neuroscience texts.
• Showcased the ability to infer brain activations in regions not explicitly reported with coordinates.
• Validated NeuroConText's effectiveness through quantitative and qualitative analyses.

Conclusions:

• NeuroConText offers a powerful solution to challenges in automated brain meta-analysis.
• The framework improves the integration of textual and spatial neuroscience data.
• Predictive brain meta-analysis can overcome limitations of sparse coordinate reporting and enhance hypothesis generation.