Author: Yang Sheng, Qingrong Jackie Wu, Qiuwen Wu, Xin Wu, Dongrong Yang 👨🔬
Affiliation: Duke University Medical Center 🌍
Purpose:
This study aims to leverage large language model (LLMs) to develop a human-in-the-loop agentic framework, enhancing the efficiency of treatment planning in radiotherapy.
Methods:
A LLM (GPT-4o) agent was developed to interact autonomously with the clinical treatment planning system (TPS) and iteratively adjust objective constraints in a human-like manner during inverse planning. The planning process was executed in a zero-shot setting, with no reliance on prior human-generated treatment plans or trajectories. To leverage the LLM's general reasoning capabilities, the expertise-dependent treatment planning task was systematically decomposed into "domain-knowledge-free" subtasks, guided by clinical context and broad planning principles. The ReAct (Reasoning + Acting) framework was employed, allowing the LLM agent to interleave reasoning traces with task-specific actions dynamically. Through this approach, the agent autonomously extracted intermediate plan statuses (e.g., dose-volume histograms, objective function loss, and dose-volume objectives), performed status analysis based on observations and historical trends using arithmetic tools, and proposed new objectives through reasoning. Following initial plan generation, human evaluation was conducted, and the agent fine-tuned the plan as necessary based on feedback to achieve optimal plan quality. Feasibility was validated using 10 head-and-neck intensity-modulated radiation therapy (IMRT) cases, with key dosimetric endpoints reported and compared to corresponding clinical plans.
Results:
The proposed workflow achieved clinical acceptability across all cases, demonstrating comparable target coverage and organs-at-risk (OARs) sparing. The median dose for the left parotid, right parotid, and oral cavity was 22.9 Gy, 20.7 Gy, and 24.5 Gy, respectively, compared to 20.7 Gy, 18.3 Gy, and 35.5 Gy in clinical plans. Conformity indices for the primary and boost PTVs were 1.60 and 1.26, compared to 1.63 and 1.59 in clinical plans.
Conclusion:
This approach represents a novel integration of LLM agents into the radiotherapy planning workflow, demonstrating their potential to enhance efficiency and decision-making in a human-centric setting.