3764 - Development and Performance Evaluation of an LLM-Based Automated Recommendation System for Radiotherapy Treatment Planning
Presenter(s)
Y. X. Yang1, L. Jia2, H. Li2, Y. Liu2, W. Zhang3, J. Zhou3, R. Guo1, X. Yu4, X. Yang1, G. Y. Wang1, G. Q. Zhou1, and Y. Sun5; 1Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China, 2Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 3Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, 4Sun Yat-sen Memorial Hospital, Guangzhou, China, 5State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China
Purpose/Objective(s): Radiotherapy (RT) planning requires meticulous optimization of dose constraints, electron density calibration, region of interest (ROI) preprocessing, and beam arrangement selection. Escalating clinical demands necessitate efficient workflows to minimize errors and enhance inter-institutional consistency. This study explores large language models (LLMs) for structured information extraction from RT planning logs to develop an automated recommendation system for personalized electron density tables, ROI preprocessing scripts, plan templates, and optimization constraints.
Materials/Methods: The study utilized anonymized operational logs from a tertiary medical center's treatment planning system (TPS). Open-source LLMs were employed for semantic parsing and knowledge extraction from unstructured log data, followed by data normalization processes including de-identification, parameter discretization, and outlier correction. Structured data were used to develop four predictive models: 1) Electron density table recommendation based on device fingerprints (manufacturer-voltage-scan region triad); 2) ROI preprocessing script prediction using tumor-type specificity and ROI topological similarity analysis; 3) Plan template selection via decision tree models integrating operator preference patterns and cancer-specific data; 4) Optimization constraint template generation combining user preference features, cancer type, and ROI similarity metrics. The final 10% of chronologically ordered data served as an independent test set to evaluate system accuracy in recommending electron density tables, ROI scripts, plan templates, and constraint templates.
Results: The LLMs successfully extracted and structured log data, with each record containing core parameters such as username, cancer type, CT parameters (manufacturer/voltage/scan region), ROI list, and electron density table, ROI preprocessing script, plan template, and optimization constraint template. The developed automated system achieved 100% accuracy in matching CT parameters with electron density tables, and similarly attained 100% accuracy for ROI preprocessing scripts in nasopharyngeal carcinoma (NPC) cases. The overall recommendation accuracy of the system reached 90% (including plan templates and optimization constraint templates).
Conclusion: This study demonstrates that LLMs exhibit exceptional semantic comprehension capabilities in structured processing of radiotherapy log data. The automated system developed based on the structured data maintained high consistency with clinical guidelines in recommending various templates, showing promising potential to reduce the need for manual intervention in clinical workflows.