Automated VMAT Planning for Cervical Cancer with and without SIB: An Optimization Strategy Driven By Large Language Model Agents
Abstract
Purpose
To evaluate a Large Language Model (LLM) for automated Volumetric Modulated Arc Therapy (VMAT) planning in cervical cancer, covering both single-target and complex Simultaneous Integrated Boost (SIB) scenarios.
Methods
70 cervical cancer patients (35 single-target, 35 SIB) treated with 6-MV photon beams on a medical linear accelerator were enrolled. Prescriptions were 36 Gy to the planning target volume (PTV) and 43 Gy to the boost volume in 20 fractions. Following In-Context Learning (ICL) on 10 cases, Qwen3-Max functioned as an autonomous agent for 60 test cases. The agent iteratively analyzed dosimetric data and adjusted optimization constraints until clinical goals were met. LLM-generated plans were compared against expert manual plans regarding Dose-Volume Histogram (DVH) indices, Conformity Index (CI), Homogeneity Index (HI), and Monitor Units (MU).
Results
Qwen3-Max generated clinically acceptable plans for all cases within an average of 99.5% (2mm / 2%).
Conclusion
LLM agents can effectively master radiotherapy planning logic through ICL. The model efficiently produces high-quality VMAT plans for complex cervical cancer cases, significantly reducing MU values and improving OAR sparing in SIB scenarios compared to expert manual planning. This framework offers a robust solution for enhancing automation and consistency in clinical treatment planning.