基于迭代更新模型的鼻咽癌自动计划质量研究

Quality improvement of automatic planning for nasopharyngeal carcinoma by iteratively updating model

目的:以鼻咽癌调强计划作为对象,应用瓦里安eclipse治疗计划系统的Rapid Plan自动计划模块,建立一种新的迭代更新方法提高Rapid Plan模型的质量,并验证该方法在临床应用的可靠性。方法:本研究随机选取鼻咽癌患者计划117例,运用计划质量评估方法及RTOG protocol 0615报告来制定计划质量评估指标对其打分。通过打分筛选每次建模的入组计划,对被筛掉的低分计划利用每次迭代更新模型辅助重新设计,并逐步提高入组计划分数,最终经过3次迭代获得最终模型。另选取20例鼻咽癌测试病例验证模型每次迭代的变化以及进行最终模型与原手工计划的比较。结果:经3次迭代分别获得模型A1、A2、A3。随着模型迭代,分数呈上升趋势,最终模型A3评分与原计划相比差异有统计学意义(P<0.05)。3个模型靶区剂量覆盖均满足临床要求,PTV_(nx)的V_(70Gy),PTV_(1)的V_(60Gy),PTV_(2)的V_(54Gy)均值差别没有超过1%。危及器官脊髓的D_(max)和视交叉的D_(max)随着模型迭代,有明显下降趋势。最终模型A3与原计划相比,脑干的V_(54Gy)、脊髓的D_(max)、左侧腮腺的D_(mean)、左右颞叶的V_(60Gy)、左右视神经以视交叉的D_(max)均下降,差异有统计学意义(P<0.05)。结论:本研究探讨了一种基于迭代更新模型的鼻咽癌自动计划方法,通过不断提升训练集质量,可以有效提高模型质量和预测能力,有效改善自动计划的计划质量,可更有效辅助物理师/剂量师设计更加优质的计划。

Objective To establish a new iteratively updating method for improving the quality of Varian eclipse Rapid Plan automatic planning model in radiotherapy planning for nasopharyngeal carcinoma(NPC),and to verify its reliability in clinical application.Methods The radiotherapy plans were randomly selected from 117 NPC patients,and the plan quality indexes were determined and scored using evaluation method and RTOG protocol 0615 report.The plans which were used for model training were screened by score threshold;and each iteratively updating model was used to assist the redesign of the low-score plans which were screened out.The score threshold was gradually improved by iterations,and the final model was obtained after 3 iterations.Another 20 NPC test cases were selected to verify the changes of the model after iteration,and the comparison between the final model and the original manual plan was carried out.Results The models A1,A2 and A3 were obtained by 3 iterations.With the iteration,the scores showed an upward trend,and the score of the final model A3 was significantly higher than the score of the original manual plan(P<0.05).The target dose coverage of the 3 models all met the clinical requirements,and their differences in the mean value of V_(70Gy)of PTV_(nx),V_(60Gy)of PTV_(1) and V_(54Gy)of PTV_(2) were not more than 1%,but the D_(max)to the spinal cord and the D_(max)to optic chiasma had an obvious downtrend with the iterations of the model.Compared with those in the original manual plan,the V_(54Gy)of brainstem,the D_(max)to spinal cord,the Dmean to left parotid gland,the V_(60Gy)of left and right temporal lobe,and the D_(max)to optic nerve and optic chiasma in the final model A3 were decreased significantly,with statistical differences(P<0.05).Conclusion The automatic planning for NPC based on iteratively updating model is explored in the study,and it is found that the model quality and prediction ability can be effectively improved through continuously improving the training set,thereby effectively enhancing the quality of automatic planning and assisting physicists or dosimetrists in designing high-quality plans.