NeuroImgAgent: a multi-agent platform for cross-species fMRI analysis with evidence-gated quality control
Cross-species functional MRI analysis requires coordinating specialized workflows and assessing intermediate outputs. We present NeuroImgAgent, a multi-agent platform driven by large language models that translates natural-language requests into workflows for humans, macaques, rats and mice. Quantitative checks and vision-language model assessments determine whether processing continues, enters bounded recovery or stops. Across 268 paired cases, preprocessing metrics were broadly comparable to those of fMRIPrep, AFNI and RABIES. With GPT-5.5, NeuroImgAgent corrected 111 of 120 induced orientation errors and passed all 52 intent-recognition and planning tasks, 8 of 9 recovery tasks and 11 of 14 tool-construction tasks on NeuroAgentBench. In 120 paired challenge cases, the full configuration increased usability from 76.7% to 92.5%, as assessed by blinded experts, and reduced unusable outputs delivered without warning from 18.3% to 4.2% of cases relative to quality-control ablation.
Authors
- Liwen Wan (ORCID: https://orcid.org/0009-0004-3340-3778)
- Yi He (ORCID: https://orcid.org/0000-0003-0040-9255)
- Hong Zhang (ORCID: https://orcid.org/0000-0002-4084-5150)
- Zeping Xie (ORCID: https://orcid.org/0000-0002-3120-5565)
- Xiaojing Long (ORCID: https://orcid.org/0000-0003-4455-7831)
- Dong Liang (ORCID: https://orcid.org/0000-0001-6257-0875)
- Jianyu Yuan
- Pingfu Wang
- Yuanxi Shen
- Yexin Han
- Zhifei Zhu
- Qian Wan
- Hairong Zheng
- Chao Zou
- Ye Li
Institutions
- Chinese Academy of Sciences (CN)
- Shenzhen Institutes of Advanced Technology (CN)
Publication Details
- Journal
- bioRxiv (Cold Spring Harbor Laboratory)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.64898/2026.11.19.767770
- Primary Topic
- Advanced MRI Techniques and Applications
- Type
- preprint