A virtual physicist framework integrating dose prediction and large language models for automated treatment planning
- 1Department of Radiation Oncology Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College Beijing China
- 2Manteia Technologies Co., Ltd Xiamen Fujian China
- 3UMass Chan Medical School Worcester Massachusetts USA
Abstract
Background: Automated radiotherapy treatment planning has benefited from data‐driven approaches; however, achieving consistently high plan quality across patients remains challenging. Dose prediction models provide patient‐specific priors but lack adaptive optimization, whereas large language models (LLMs) offer strong reasoning capabilities but raise concerns regarding stability and safety in clinical use.
Purpose: To develop and validate a fully automated VMAT treatment planning framework integrating dose prediction and LLM‐driven dynamic optimization, termed the Virtual physicist (VP), and to evaluate its dosimetric performance, plan complexity, and consistency relative to manual and other automated planning strategies.
Methods: A total of 70 cervical cancer patients previously treated with VMAT were retrospectively collected. The dataset was divided into 50 cases for training, 10 for validation, and 10 for testing. The input data consisted of patients’ planning CT images with contoured planning target volumes (PTVs) and organs at risk (OARs). The outputs included predicted three‐dimensional dose distributions (upper‐, middle‐, and lower‐dose bands) and automatically generated VMAT treatment plans with corresponding dosimetric and plan complexity metrics. The VP framework was implemented in the MOZI treatment planning system via a Python scripting interface. A dose‐band prediction model was used to initialize patient‐specific optimization constraints. During optimization, an LLM iteratively adjusted objectives and weights based on real‐time dosimetric feedback, subject to a rule‐based safety engine enforcing clinical and physical constraints. Plans generated using dose prediction only (DP‐only), LLM only (LLM‐only), and combined dose prediction with LLM (DP + LLM) were compared with clinically delivered manual plans.
Results: The DP + LLM strategy significantly improved sparing of the bladder and bilateral femoral heads compared with all three other strategies and significantly reduced rectal doses relative to the DP‐only and manual plans (P < 0.05), while maintaining PTV Dmean, HI, CI, and GI comparable to those of the manual plans (P > 0.05). For dual‐layer MLC plan complexity, DP + LLM produced a more balanced modulation pattern than manual planning, with lower distal‐layer complexity—significant for d‐C/A and d‐MCS but not d‐BA—and higher proximal‐layer complexity across all three metrics. Compared with DP‐only and LLM‐only plans, DP + LLM showed a consistent trend toward greater complexity in most metrics, with significant differences in several metrics. In addition, the automated strategies generally showed smaller standard deviations in complexity metrics, indicating improved planning consistency.
Conclusion: Integrating dose prediction with LLM‐driven dynamic optimization enables a robust and fully automated planning workflow, improving OAR sparing and planning consistency without compromising target dose quality, and demonstrates strong potential for clinically applicable intelligent radiotherapy planning.
Related
- Generalized multi-energy model for computing radiation treatment parameters on image domain and automated model-selection framework
- Niels Bohr as a Physicist of Principle
- A novel deep learning prediction method for PM2.5 integrating improved MSTL decomposition and FATA optimization algorithm
- kicad-pycbnew: A Python Framework for Automated, High-Precision PCB Design in Particle Detector Instrumentation
- Automated Many-Body Simulations of Strongly Correlated Systems Using a Correlation-Aware Agentic Framework
BibTeX
@article{zhu2026virtual,
title = {{A virtual physicist framework integrating dose prediction and large language models for automated treatment planning}},
author = {Heling Zhu and Junqiang Li and Yongguang Liang and Wenbo Li and Yiming Zhang and Wei Zhang and Bo Yang and Qichao Zhou},
journal = {Journal of Applied Clinical Medical Physics},
year = {2026},
month = nov,
doi = {10.1002/acm2.70904},
url = {https://doi.org/10.1002/acm2.70904},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A virtual physicist framework integrating dose prediction and large language models for automated treatment planning
AU - Heling Zhu
AU - Junqiang Li
AU - Yongguang Liang
AU - Wenbo Li
AU - Yiming Zhang
AU - Wei Zhang
AU - Bo Yang
AU - Qichao Zhou
PY - 2026
DA - 2026/11/05
JO - Journal of Applied Clinical Medical Physics
DO - 10.1002/acm2.70904
UR - https://doi.org/10.1002/acm2.70904
AB - Background: Automated radiotherapy treatment planning has benefited from data‐driven approaches; however, achieving consistently high plan quality across patients remains challenging. Dose prediction models provide patient‐specific priors but lack adaptive optimization, whereas large language models (LLMs) offer strong reasoning capabilities but raise concerns regarding stability and safety in clinical use. Purpose: To develop and validate a fully automated VMAT treatment planning framework integrating dose prediction and LLM‐driven dynamic optimization, termed the Virtual physicist (VP), and to evaluate its dosimetric performance, plan complexity, and consistency relative to manual and other automated planning strategies. Methods: A total of 70 cervical cancer patients previously treated with VMAT were retrospectively collected. The dataset was divided into 50 cases for training, 10 for validation, and 10 for testing. The input data consisted of patients’ planning CT images with contoured planning target volumes (PTVs) and organs at risk (OARs). The outputs included predicted three‐dimensional dose distributions (upper‐, middle‐, and lower‐dose bands) and automatically generated VMAT treatment plans with corresponding dosimetric and plan complexity metrics. The VP framework was implemented in the MOZI treatment planning system via a Python scripting interface. A dose‐band prediction model was used to initialize patient‐specific optimization constraints. During optimization, an LLM iteratively adjusted objectives and weights based on real‐time dosimetric feedback, subject to a rule‐based safety engine enforcing clinical and physical constraints. Plans generated using dose prediction only (DP‐only), LLM only (LLM‐only), and combined dose prediction with LLM (DP + LLM) were compared with clinically delivered manual plans. Results: The DP + LLM strategy significantly improved sparing of the bladder and bilateral femoral heads compared with all three other strategies and significantly reduced rectal doses relative to the DP‐only and manual plans (P < 0.05), while maintaining PTV Dmean, HI, CI, and GI comparable to those of the manual plans (P > 0.05). For dual‐layer MLC plan complexity, DP + LLM produced a more balanced modulation pattern than manual planning, with lower distal‐layer complexity—significant for d‐C/A and d‐MCS but not d‐BA—and higher proximal‐layer complexity across all three metrics. Compared with DP‐only and LLM‐only plans, DP + LLM showed a consistent trend toward greater complexity in most metrics, with significant differences in several metrics. In addition, the automated strategies generally showed smaller standard deviations in complexity metrics, indicating improved planning consistency. Conclusion: Integrating dose prediction with LLM‐driven dynamic optimization enables a robust and fully automated planning workflow, improving OAR sparing and planning consistency without compromising target dose quality, and demonstrates strong potential for clinically applicable intelligent radiotherapy planning.
SN - 1526-9914
ER -