NeuroPilot AI agents slash neuroimage processing from months to 1 week

NeuroPilot, a new multi-agent system described in a July 2026 arXiv paper (arXiv:2608.07541), automates the entire neuroimaging pipeline—from raw DICOM files to analysis-ready derivatives. Built by a team including Yiyao Chen and Guorong Wu, the system digitizes expert knowledge into three LLM-invocable skills: dcm2bids-skill for data standardization, neuroimage-pre-skill for modality-specific preprocessing, and qc-agent-skill for evidence-based quality control. The LLM-driven agent dynamically routes datasets to optimal pipelines based on cohort traits, e.g., dispatching T1w and fMRI data to fMRIPrep or selecting specialized processing for infant brains. This design eliminates project-specific scripts and environment-adaptive tuning, requiring only a single configuration.

NeuroPilot's agent then drives a semi-automated QC workflow using a 3-D browser dashboard with a multi-tiered verification system, automatically optimizing failed cases and escalating complex issues to human supervisors. In deployment across 17 cohorts (123,000+ subjects spanning infant to aging populations and multiple MRI modalities), the QC agent screened 558 production subjects, with automated flags validated against FreeSurfer's topology-defect metrics. The infant processing pipeline achieved a 100% completion rate (201/201) on QC-validated inputs. Most notably, NeuroPilot compresses the traditional 2–3 month timeline for training staff and processing complete datasets into a single week—a dramatic efficiency gain that could accelerate large-scale neuroimaging research and clinical translation.