MARTP: a multi-agent simulation framework for automated radiation therapy planning based on LLMs.

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiotherapy Planning

Background:

• Current radiotherapy planning is labor-intensive, time-consuming, and lacks scalability.
• There is a need for intelligent automation to improve precision and efficiency in radiotherapy.
• Multidisciplinary collaboration is crucial but challenging to integrate into clinical workflows.

Purpose of the Study:

• To propose MARTP, a Multi-Agent Radiation Therapy Planning framework driven by large language models (LLMs).
• To emulate multidisciplinary clinical workflows for end-to-end intelligent radiotherapy planning and evaluation.
• To enhance precision, efficiency, and scalability in radiotherapy planning.

Main Methods:

• Developed a Multi-Agent Radiation Therapy Planning (MARTP) framework using LLMs.
• Integrated five specialized agents for data analysis, weight adjustment, optimization, evaluation, and reporting.
• Employed supervised fine-tuning (SFT), retrieval-augmented generation (RAG), and reinforcement learning (RL) for plan generation and evaluation.

Main Results:

• MARTP generated plans with dosimetric metrics comparable to expert-crafted plans.
• Significantly improved radiotherapy planning efficiency was observed.
• The framework demonstrated robust and safe performance, even with abnormal inputs.

Conclusions:

• LLM-driven multi-agent systems can effectively replicate radiotherapy workflows.
• MARTP offers a promising approach for intelligent automation in radiotherapy.
• The framework supports consistent decision-making and scalable treatment planning.