169 - Automating Head-and-Neck Cancer Intensity Modulated Radiation Therapy Treatment Planning with a ReAct Large Language Model Agent

Purpose/Objective(s):

Large language models (LLMs) have recently demonstrated exceptional capabilities, offering the potential to streamline workflows and enhance efficiency across diverse tasks. However, their application in domain-specific areas, such as radiotherapy treatment planning, remains challenging due to the lack of publicly available, specialized knowledge required for these tasks. In this study, we aim to address these limitations by augmenting an LLM with custom-designed functions and implement the LLM to perform treatment planning for head-and-neck cancer.

Materials/Methods:

In this study, 10 head-and-neck cancer patients who received Intensity-Modulated Radiation Therapy (IMRT) at our institution were retrospectively collected under IRB approval. A LLM agent was implemented to directly interact with the clinical treatment planning system (TPS) to iteratively adjust inverse optimization constraints. Following the ReAct (Reasoning + Acting) framework, the agent iteratively extracts intermediate plan states and proposes new constraint values to guide inverse optimization. To enhance the agent’s arithmetic reasoning and streamline its interaction with the TPS, supporting functions were developed for dosimetric deviation computation, gradient calculation and TPS interaction. Each interaction cycle consists of three sequential submodules: load state, analyze state, and update constraints. Within each module, the agent observes, reasons, and acts using specialized tools for interaction or analysis. The agent’s decision-making process is informed not only by current observations but also by an adaptive memory module that records previous optimization attempts and evaluations, allowing for dynamic strategy refinement. The planning process was performed in a zero-shot inference setting, where the LLM operates without prior exposure to manually generated treatment plans and is utilized without any fine-tuning or task-specific training. The LLM-generated plans were evaluated against clinical plans, with key dosimetric endpoint statistics analyzed and reported.

Results:

The LLM-generated plans demonstrated comparable organs-at-risk (OARs) sparing to clinical plans while achieving improved hot spot control (Dmax: 106.2% vs. 108.0%, p<0.05; homogeneity index: 0.059 vs. 0.061, p=0.55) and better conformity (conformity index: 1.26 vs. 1.59, p<0.05).

Conclusion:

This study customizes and implements an LLM agent in the radiation therapy treatment planning, demonstrating a new philosophy of automating treatment planning. By equipping the LLM with tailored tools and harnessing its general reasoning abilities, we enabled it to autonomously refine constraint adjustments during the inverse planning process, improving its ability to navigate complex treatment planning decisions.