Automated radiotherapy treatment planning guided by GPT-4Vision

Area of Science:

• Radiation Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Radiotherapy treatment planning is complex, time-consuming, and subjective.
• Advanced Artificial Intelligence (AI) models offer potential solutions for planning challenges.
• Integrating AI with clinical knowledge can enhance decision-making.

Purpose of the Study:

• Introduce GPT-RadPlan, an automated framework using GPT-4 Vision for radiotherapy planning.
• Leverage multi-modal AI to integrate radiation oncology knowledge and reasoning capabilities.
• Automate treatment planning to improve efficiency and outcomes.

Main Methods:

• Utilize in-context learning with clinical requirements and approved plans for GPT-4 V.
• Integrate GPT-RadPlan into an existing inverse treatment planning system via API.
• Employ GPT-RadPlan as both plan evaluator and planner, iteratively refining plans through textual feedback.

Main Results:

• GPT-RadPlan was tested on 17 prostate and 13 head & neck cancer VMAT plans.
• The system matched or outperformed human expert plans in all cases.
• Achieved superior target coverage and reduced organ-at-risk doses by an average of 5 Gy (10-15%).

Conclusions:

• GPT-RadPlan is the first multimodal large language model agent for automated radiotherapy planning.
• The system successfully mimics human planner behavior and adheres to clinical protocols.
• Demonstrates significant promise in automating radiation oncology treatment planning without retraining.