Author: Peng Dong, Elizabeth Kidd, Sheng Liu, Thomas R. Niedermayr, Oscar Pastor-Serrano, Lei Xing, Yong Yang, James Zou 👨🔬
Affiliation: Department of Biomedical Data Science, Stanford University, Department of Radiation Oncology, Stanford University, Stanford University 🌍
Purpose: Radiotherapy treatment planning is a time-consuming and potentially subjective process that requires iterative adjustments of optimization parameters to balance conflicting objectives. In this study, we aim at developing an automated treatment planning plugin that integrates with the clinically widely used Eclipse Treatment Planning System (TPS). Using a large language model (LLM) backbone, the resulting GPT-RadPlan plugin acts as an expert planner capable of guiding the planning process by adjusting optimization weights, objective doses, and dynamically adding or removing planning objectives.
Methods: GPT-RadPlan iteratively evaluates treatment plans and suggests new optimization settings based on its own suggestions. At the beginning of each iteration, GPT-RadPlan’s evaluation module assesses dose distributions and dose-volume histograms (DVHs), providing detailed textual feedback on plan quality and level of agreement with the physician’s intent. Based on (i) the evaluation module’s feedback, (ii) the parameters of previous iterations, and (iii) the optimization settings from three previous approved plans for the same disease site and prescription; the planner module adjusts optimization weights, objective doses, and planning objectives to meet clinical requirements.
Results: GPT-RadPlan-generated plans were compared to 8 clinically approved plans for VMAT prostate cancer treatments. GPT-RadPlan either outperformed or matched the clinical plans, similarly covering the targets while reducing organ-at-risk (OAR) doses 11% on average. Specifically, GPT-RadPlan achieved an average bladder and rectum mean dose of 26.9 Gy and 30.23 Gy, compared to 29.95 Gy and 35.86 Gy for the clinical plans, respectively. The plugin consistently satisfied all dosimetric objectives outlined in the physician’s intent, demonstrating its ability to adjust planning parameters dynamically and optimize plan quality.
Conclusion: We introduce an LLM-powered plugin that mimics expert human planners within a commercial TPS like Eclipse. By automating the planning process, including adjusting optimization parameters and modifying planning objectives, GPT-RadPlan streamlines workflow and can potentially standardize plan quality.