Teaching an LLM to Learn: A Self-Learning Approach for Autonomous Radiotherapy Planningcopilot for Locally Advanced Lung Cancer
Abstract
Purpose
To evaluate whether a Large Language Model (LLM)–driven autonomous planning system can self-learn planning strategies from human planner logs and apply this knowledge to generate clinically compatible radiotherapy plans without manual refinements.
Methods
The PlanningCopilot is an LLM-driven autonomous radiotherapy planning system that understands planning goals, selects optimization actions, and iteratively refines plans through the rest API to execute a set of plan optimization actions. Planner logs from 10 stage II-IIIA locally advanced non-small cell lung cancer (LA-NSCLC) patients were used as a strategy-learning cohort to capture each optimization step, including the selected action, parameters, and resulting DVH metrics. An LLM (GPT-5) analyzed these logs to extract heuristic relationships between optimization actions and DVH responses, producing a concise planning guideline outlining action selection criteria and parameter ranges. This self-learned guideline was subsequently used by PlanningCopilot to autonomously generate plans for 59 LA‑NSCLC from a separate evaluation cohort patients without any manual plan refinements. Plan quality was evaluated by comparing autonomous plans with corresponding clinical plans using DVH metrics. Statistical comparisons were performed using Wilcoxon signed‑rank tests.
Results
PlanningCopilot successfully generated clinically compatible plans for all 59 patients using only the self-learned planning guidelines. No statistically significant differences were observed between autonomous and clinical plans for PTV D0.03cc (p=0.50), Cord Dmax (p=0.56), Lungs V20Gy (p=0.49), Lungs Dmean (p=0.48), or Esophagus Dmean (p = 0.99). A modest but statistically significant reduction in esophagus Dmax was observed in autonomous plans (p = 0.040).
Conclusion
This proof-of-concept study demonstrates that an LLM can extract planning strategies from limited human planner logs and apply them to autonomously generate LA-NSCLC treatment plans with compatible plan quality relative to clinical plans. While promising, further validation using larger and multi-institutional datasets is required to assess generalizability and clinical robustness.