Radiation treatment planning must balance competing objectives to meet clinical criteria while respecting physicians' preferences. This paper introduces an evidence‑driven multi‑agent compound AI system that combines deep‑reinforcement‑learning (DRL) based automated planning with large‑language‑model (LLM) conversational refinement, targeting proton pencil‑beam scanning (PBS) therapy.
First, a policy network trained with proximal policy optimization (PPO) generates multiple plans per patient without human intervention. One plan is selected as the target, while the others serve as references for subsequent refinements. For challenging cases that require further adjustments, an LLM‑powered multi‑agent system enables plan refinement through natural‑language interaction. The system comprises:
- a planning agent based on Qwen3‑14B that proposes evidence‑backed adjustments;
- a chatbot offering a unified conversational interface;
- an evaluation agent that retrieves clinical metrics and flags constraint violations.
To enhance domain‑specific reasoning, the planning agent is post‑trained on task‑specific data using supervised fine‑tuning (SFT) and gradient‑reward‑based policy optimization (GRPO). The framework was trained and evaluated on 21 head‑and‑neck cancer patients, each with five PPO‑generated plans (105 plans total), of which 50 were selected for instruction‑guided refinement. Results show that the post‑trained Qwen3‑14B outperforms the base LLM in instruction following, information retrieval, and domain reasoning. The refined adjustments more effectively improve user‑prioritized objectives while limiting deterioration of competing objectives, yielding clinically appropriate patient‑specific trade‑offs.
The proposed compound AI system offers a user‑friendly planning platform and presents a promising approach to personalized proton PBS treatment planning.
Review