Fully Automated Robust Intensity Modulated Proton Therapy Planning for Prostate Cancer with and without Pelvic Lymph Node Irradiation Using a Frontier Large Language Model
Abstract
Purpose
To develop and validate an auto-planning approach for robust intensity modulated proton therapy (IMPT) using a frontier large language model (LLM), ChatGPT-4o, in a clinical treatment planning system (TPS) for prostate cancer, automatically managing heterogeneous prescriptions, protocols, and contours, including patients with or without pelvic lymph node irradiation (PLNI), using single field optimization (SFO) or multi-field optimization (MFO).
Methods
A multi-agent LLM-driven auto-planning approach was developed, including a secure interface between ChatGPT-4o and our clinical TPS (RayStation 2023B). The auto-planning inputs included a contoured CT, prescription, clinical goals (CGs), and a free-text description of patient-specific priorities. The LLM executes all subsequent planning steps, including prescription definition, isocenter and beam placement, SFO vs. MFO selection, planning structure and optimization objective creation, and iterative robust optimizer execution and objective tuning. The approach was tested in 13 retrospective prostate cancer patients, including 4 and 9 with and without PLNI, respectively, and the CG pass rate was compared to clinical plans produced by expert human planners.
Results
The LLM approach achieved a mean±standard deviation (SD) CG pass rate of 94.4% ± 3.3%, nearly identical to the 94.4% ± 3.1% for the clinical plans, and correctly applied SFO or MFO strategies in every case. The LLM approach achieved more, the same, or fewer CGs in 7, 2, and 4 cases, respectively. The mean±SD execution time was 190.6±20.8 min and 46.8±15.7 min for cases with and without PLNI, respectively.
Conclusion
The LLM approach successfully generated highly complex, multi-phase robust IMPT plans while meeting safety objectives at the same rate as expert humans in the setting of true clinical contouring and treatment directive variations. The approach holds promise to significantly improve the speed, consistency, and quality of IMPT planning, and we are currently pursuing clinical implementation.