NEURA: An agentic system for autonomous neuroimaging workflows
Journal:
bioRxiv
Published Date:
Apr 30, 2026
Abstract
Neuroimaging is essential for studying the human brain; however, the deep interdisciplinary expertise required imposes a very high threshold, limiting its broader clinical and scientific applications. We introduce NEURA, a large language model (LLM)-powered agentic system for automated neuroimaging workflow planning and analysis. NEURA processes free-text research questions and multimodal neuroimaging datasets to generate evidence-grounded analysis plans, executable scripts, validated statistical results and structured reports, with traceable reasoning linked to intermediate artefacts and full execution records. Through extensive evaluations on a curated benchmark, NEURA achieved an 89.5% planning accuracy and substantially outperformed direct LLM queries, with average gains of 30.5% in planning accuracy, 25.6% in tool selection and 36.7% in tool ordering. In case studies of spinocerebellar ataxia type 3, NEURA autonomously identified cerebellar atrophy and abnormal diffusivity patterns consistent with established pathologies and expert manual analyses. Collectively, these results demonstrate that our work advances from pipeline automation to rigorous, scalable and interpretable neuroimaging research systems.
