Title: Developing a Virtual Neuroscientist: Autonomous Neuroimaging Analysis Through Multi-Agent Collaboration

Abstract: Converting neuroimaging data into clinically useful biomarkers is a process that demands significant domain expertise and manual effort. While standardized pipelines like fMRIPrep have enhanced efficiency and robustness, they operate on static configurations. Consequently, they lack the ability to reason about downstream goals, evaluate alternative strategies, or iteratively adjust intermediate steps based on new evidence—capabilities inherent to human researchers. This absence of closed-loop adaptation frequently forces experts into tedious, manual cycles of parameter tuning and error correction, thereby limiting the scalability of clinical biomarker development. To address these limitations, we present NEXUS, an autonomous multi-agent framework that combines neuroimaging workflow execution with a deep understanding of scientific objectives. Moving beyond traditional flat tool-calling agents, NEXUS utilizes a code-centric execution model in which specialized agents collaboratively build and optimize executable programs using composable, domain-specific primitives. This architecture supports the creation of resilient, long-horizon workflows that dynamically adapt to runtime observations. Additionally, we introduce a hierarchical verification system for autonomous quality control, which merges cohort-level metric screening with agentic visual inspection to facilitate evidence-based workflow corrections. Evaluations on the ADHD-200 and ADNI datasets show that NEXUS surpasses standard workflow-based baselines in predictive accuracy while demonstrating complex agentic behaviors, such as strategy exploration and adaptive refinement. The source code is accessible at https://github.com/LearningKeqi/Virtual-Neuroscientist-NEXUS.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC