Reference-Free large language model agents for physician-guided radiotherapy treatment planning

Area of Science:

• Artificial Intelligence in Medicine
• Radiation Oncology
• Medical Physics

Background:

• Large language models (LLMs) show promise for task efficiency but face challenges in specialized domains like radiotherapy.
• Adapting LLMs for clinical radiotherapy planning requires domain-specific knowledge, which is often not publicly available.

Purpose of the Study:

• To adapt a general large language model (LLM) into a clinically applicable agent for intensity-modulated radiation therapy (IMRT) inverse treatment planning.
• To investigate the essential architectural and functional components for creating a viable LLM-based planning agent.

Main Methods:

• A large language model (LLM) agent was developed to interact with a treatment planning system (TPS) for head and neck cancer patients.
• The LLM agent iteratively adjusted optimization constraints based on real-time plan evaluations in a reference-free setting.
• Plans generated by the LLM were compared to those created by certified dosimetrists using key dosimetric endpoints.

Main Results:

• LLM-generated plans demonstrated comparable organ-at-risk (OAR) sparing to clinical plans.
• Improved hot spot control (Dmax: 106.5% vs. 108.8%, p < 0.05) and superior conformity (CI: 1.18 vs. 1.39, p < 0.05 for boost PTV) were observed.
• The LLM operated effectively without prior exposure to manual treatment plans or fine-tuning.

Conclusions:

• A reference-free, LLM-driven workflow for automated IMRT treatment planning in a commercial TPS is feasible.
• This approach offers a generalizable solution to reduce planning variability in radiation therapy.
• The study supports the broader adoption of AI-based planning strategies in clinical practice.