Split-Brain Auto-Planning a Hierarchical Multi-Agent LLM Framework for Inverse Optimization
Abstract
Purpose
To demonstrate a novel "Split-Brain" Multi-Agent Large Language Model framework for autonomous inverse planning in radiation oncology.
Methods
We developed a dual-agent architecture within an open-source inverse planning framework (PortPy). A Reasoning Agent (DeepSeek-R1, Attending Physicist role) reviews dose volume histogram summaries and spatial dosimetric metrics to generate high-level clinical directives, such as prioritizing PTV coverage versus hotspot control. An Action Agent (Qwen-2.5, Planner role) converts these directives into explicit updates to objective function weights and penalty terms, then re-runs optimization iteratively. We evaluated the system on multiple lung cancer cases initialized with suboptimal objective weights. Performance was assessed by improvements in PTV D95%, hotspot control, and maintenance of organ-at-risk doses within tolerance limits.
Results
Across lung cases, the framework consistently improved plans starting from intentionally suboptimal weights, reaching prescription-level target coverage while strictly maintaining critical organ (e.g., Spinal Cord, Heart) constraints. The system demonstrated autonomous clinical reasoning by shifting focus to hotspot control when doses exceeded a clinical threshold (115% of prescription). In a representative case, the framework dramatically improved PTV D95% within a few iterations and subsequently reduced hotspots by autonomously escalating overdose penalties. Once clinical goals were met, plan metrics showed no further meaningful change with additional iterations, indicating stable convergence without continued retuning.
Conclusion
The Split-Brain LLM framework demonstrates the feasibility of autonomous clinical reasoning for treatment planning. By framing decisions as explicit strategies instead of opaque “black-box” behaviors, this approach provides a transparent, robust path to next-generation automated treatment planning.