2157 - Expanding GPT-RadPlan: An Agentic Framework for Fully Automated and Adaptive Radiotherapy Treatment Planning
Presenter(s)
S. Liu1, S. Wang2, P. Dong3, Y. Yang2, J. Zou1, and L. Xing2; 1Department of Biomedical Data Science, Stanford University, Stanford, CA, 2Department of Radiation Oncology, Stanford University, Stanford, CA, 3Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA
Purpose/Objective(s): Radiotherapy treatment planning is labor-intensive, requiring iterative optimization to balance conflicting clinical objectives. Building on GPT-RadPlan, an LLM-powered solution integrated with a treatment planning system, this study introduces an Agentic Framework for higher automation and adaptability. Unlike traditional fixed workflows, this framework allows autonomous agents to independently decide when to evaluate, predict optimization weights, and select tools, aiming to deliver consistent, high-quality plans across diverse anatomical sites, including complex head and neck (H&N) regions, while reducing planning time and variability.
Materials/Methods: The enhanced GPT-RadPlan was tested on external beam radiotherapy cases of varying complexity, including H&N, lung, and prostate. The framework employs autonomous agents for plan evaluation, optimization, and tool selection. Agents dynamically determine evaluation timing and choose tools based on task complexity, using high-resolution image analysis for intricate cases and rapid DVH generation for simpler ones. Actor agents adapt optimization by predicting weights, modifying objectives, and introducing synthetic organs-at-risk (OARs) when needed. A central orchestrator agent coordinates real-time interactions, prioritizing decisions based on clinical requirements. The process runs autonomously until clinical goals are met or improvements plateau. Statistical analyses assessed efficiency, target coverage, and OAR sparing to evaluate scalability, robustness, and generalizability.
Results: Preliminary evaluations show that the Agentic GPT-RadPlan generalizes complex external beam cases, including H&N. The agent-driven approach produced clinically acceptable plans with improved efficiency, demonstrating potential benefits in planning speed and quality. Agents effectively adjusted optimization parameters, applied adaptive strategies like synthetic OAR generation, and selected tools based on case complexity, ensuring consistent plan quality across diverse cases.
Conclusion: This study highlights the transformative potential of an Agentic Framework powered by GPT-RadPlan in fully automating radiotherapy planning. By enabling autonomous agents to independently manage evaluation timing, optimization adjustments, and tool selection, the framework achieves superior automation and adaptability compared to conventional workflows. Dynamic agent coordination ensures flexible, responsive planning optimized for each case’s challenges. The framework offers a scalable, efficient, and clinically robust solution, minimizing human intervention and delivering standardized, high-quality plans across diverse disease sites.