摘要目的:研究利用多目标优化（MCO）技术提高基于先验知识自动计划（KBP）模型的计划设计质量的可行性。方法:选择55例已完成放疗的鼻咽癌患者，每个病例都采用固定野调强放疗技术。随机选择40个病例的调强放疗计划作为训练集1，通过多目标优化技术对训练集1中的放疗计划进行预处理，构建新的训练集2。将初始训练集1和处理后的训练集2作为样本，分别训练得到传统的KBP模型和多目标优化技术精炼后的MCO-KBP模型。在剩余15个病例中随机选择5例作为验证集，其余10例作为测试集。验证通过后用测试集对初始的人工计划以及传统KBP模型和MCO-KBP模型生成自动计划的计划质量进行统计学分析。结果:传统KBP模型和MCO-KBP模型生成自动计划的靶区剂量（D 95%）均能满足临床要求，适形性指数与均匀性指数基本一致（ P>0.05），且MCO-KBP模型生成的自动计划中重要危及器官的剂量比传统KBP模型更低（脑干D max平均偏低2.13 Gy，左、右腮腺D mean分别平均偏低1.39、1.59 Gy，左、右视神经及垂体的D max分别平均偏低1.42、1.16、1.88 Gy，差异均有统计学意义）。 结论:与传统的KBP模型相比，精修后的MCO-KBP模型设计的调强放疗计划对危及器官的保护具有明显优势，利用MCO技术提高KBP模型的计划设计质量是可行的。

abstractsObjective:Utilizing multi-criterion optimization (MCO) technology to improve plan design quality based on knowledge-based planning (KBP) model.Methods:Fifty-five patients with nasopharyngeal carcinoma (NPC) who had completed radiotherapy were selected, and fixed-field intensity-modulated radiotherapy (IMRT) technology was used in each case. Among them, 40 cases were randomly selected as training set 1. Then, IMRT plans in training set 1 were preprocessed by MCO technology to construct a new training set 2. With the initial training set 1 and the processed training set 2 as training samples, the traditional KBP model and the MCO-KBP model refined by MCO technology were trained, respectively. Among the remaining 15 cases, 5 cases were randomly selected as the validation set, and the remaining 10 cases were used as the test set. After verification, the test set was used to statistically analyze the plan quality of the initial manual plan and the automatic plan generated by the traditional KBP model and the MCO-KBP model.Results:The target dose (D 95%) of plans generated by the traditional KBP model and the MCO-KBP model met the clinical requirements. Conformity index (CI) and homogeneity index (HI) were almost the same ( P>0.05), and the doses of organ at risk (OAR) of the automatic plans generated by the MCO-KBP model were lower than those of the traditional KBP model. For example, compared with the traditional KBP model, the average D max of the brainstem in the automatic plans generated by the MCO-KBP model was lower by 2.13 Gy, the average D mean of the left parotid gland was lower by 1.39 Gy, the average D mean of the right parotid gland was lower by 1.59 Gy, and the average D max of the left optic nerve was lower by 1.42 Gy, the average D max of the right optic nerve was lower by 1.16 Gy, and the average D max of the pituitary gland was lower by 1.88 Gy. All of the above-mentioned dosimetry indexes were statistically significant. Conclusion:Compared with the traditional KBP model, the IMRT plans designed by the refined MCO-KBP model have obvious advantages in the protection of OAR, which proves the feasibility of utilizing MCO technology to improve the plan design quality of the KBP model.