Towards Intelligent Agents for Radiotherapy: Integrating Exploration-Exploitation with Foundation Models

Area of Science:

• Medical Physics
• Artificial Intelligence in Oncology
• Radiotherapy Treatment Planning

Background:

• Radiotherapy treatment planning involves complex optimization problems, such as Beam Angle Optimization (BAO).
• Manual planning is labor-intensive and requires significant expertise from medical physicists.
• Existing automated methods may not fully capture the nuances of dose distribution analysis.

Purpose of the Study:

• To develop an automated approach for radiotherapy treatment planning using a multimodal Large Language Model (LLM).
• To investigate the efficacy of an LLM-based iterative framework for Beam Angle Optimization (BAO).
• To enhance treatment plan quality by improving dose conformity and organ-at-risk sparing.

Main Methods:

• Integration of a reinforcement-learning-style iterative framework with a multimodal LLM (GPT-4V).
• Utilizing GPT-4V for candidate beam angle selection and analysis of Monte Carlo-simulated dose distributions (MatRAD).
• Employing a reward function for iterative plan refinement and exploration-exploitation principles.

Main Results:

• The LLM-based framework demonstrated superior performance in Beam Angle Optimization compared to random selection.
• The proposed method outperformed deep reinforcement learning baselines in treatment plan quality.
• Experimental results on prostate cancer cases show promising outcomes for LLM-assisted planning.

Conclusions:

• Large Language Models show significant potential for assisting in complex radiotherapy treatment planning tasks.
• The automated approach can alleviate manual planning efforts and improve plan quality.
• This LLM-based system aids medical physicists in exploring beam configurations and refining plans for better outcomes.