摘要目的:开发并评估一种针对治疗计划系统（TPS）的调强放疗（IMRT）无监督自动计划方案，使其能够模拟人工进行治疗计划的自动优化。方法:回顾性分析2022年3月至2023年3月浙江省肿瘤医院已经完成放疗的25例胃癌患者资料，患者年龄40~60岁，其中训练集7例，测试集18例。所有患者均采用相同的临床处方剂量标准45 Gy分25次，并接受飞利浦大孔径腹部CT扫描，扫描层厚为5 mm。基于深度强化学习（DRL）框架，提出一种多智能体优化决策网络（MOPN），对多个优化目标进行调整，从而模拟临床人工计划设计的过程。所有病例的自动计划方案均借助Eclipse脚本应用程序接口（ESAPI）进行代码编程，由MOPN模型自动生成。利用Wilcoxon符号秩检验比较自动计划方案与人工计划方案在相关剂量学指标间的差异。结果:初始优化目标经过MOPN调整后，自动计划的平均得分由（576.1±221.2）分上升至（1852.8±294.9）分。与临床人工计划相比，MOPN自动计划在脊髓D max、肝D mean和肝V 5 Gy方面分别降低了21.4%、9.8%和11.5%。 结论:MOPN模型借助ESAPI工具完成了与TPS的数据互通，同时也实现了胃癌IMRT治疗计划的自动化设计。经过训练的MOPN模型可以模仿计划者在优化过程中的人为操作来调整多个目标，逐步改善计划质量。

abstractsObjective:To develop and evaluate an unsupervised intensity-modulated radiation therapy (IMRT) automated planning scheme for the Eclipse commercial treatment planning system (TPS), aiming to simulate the manual operation during the whole optimization process.Methods:A retrospective analysis was performed on 25 gastric cancer patients aged 40-60 years who had completed radiotherapy in Zhejiang Cancer Hospital from March 2022 to March 2023. All patients were divided into the training ( n=7) and test sets ( n=18). All patients were treated with the same clinically prescribed dose standard: 45 Gy/25 times. Abdominal CT scan was performed using Philips simulator with a thickness of 5 mm. Based on the deep reinforcement learning (DRL) framework, a multi-agent optimization policy network (MOPN) was proposed to simulate the process of clinical manual planning design and obtain high quality automatic planning according to adjusting multiple optimization objectives. The automatic plan for all cases was generated by code programming using the eclipse scripting application program interface (ESAPI). Wilcoxon signed rank test was used to investigate the significance of the difference between automatic planning and clinical manual planning. Results:After the initial optimization objectives were adjusted by MOPN, the average plan score of all automatic plans was increased from 576.1±221.2 to 1852.8±294.9. Compared with clinical manual plans, the average D max of the spinal cord, the average D mean and V 5 Gy of the liver in the MOPN plans were reduced by 21.4%, 9.8% and 11.5%, respectively. Conclusions:With the help of ESAPI tool, MOPN can realize data interaction with TPS and the automation of IMRT treatment plan for gastric cancer. The trained MOPN can mimic the manual operation of the planner to adjust multiple optimization objectives and gradually improve the plan quality.