探讨NSCLC放射性肺损伤NTCP模型的建立

Establishment of the NTCP model for radiation-induced lung injury in patients with non-small cell lungcancer

目的 针对局部晚期NSCLC受累野IMRT后RILI发生情况,分析不同模型建模方法并评价其预测效果。方法 回顾分析2007—2011年间肿瘤医院收治的242例未进行手术治疗的Ⅲ期NSCLC患者临床资料。以放疗结束后6个月内发生2、3级RILI分别作为终点事件。分别采用PCA模型、LKB模型和MLD模型3种计算方法,对剂量学参数建立NTCP预测模型,并评价其预测效果。结果 PCA模型提取4个主成分,2、3级RILI预测结果,AUC分别为0.650、0.606。LKB模型2级RILI参数拟合得到m=0.46,n=1.35,D50=23.59 Gy,3级RILI参数拟合得到m=0.36,n=0.27,D50=72.67 Gy,AUC分别为0.607、0.585。MLD模型2、3级参数预测结果γ50=1.073,D50=24.66 Gy和γ50=0.97,D50=48.45 Gy,AUC分别为0.604和0.569。结论 使用相同治疗模式的单一人群大样本数据进行建模,对提高模型预测准确性和稳定性很重要。LKB模型和PCA模型均能较好地预测RILI的发生概率,MLD模型对3级RILI的预测效果较差。

Objective To examine the incidence of radiation-induced lung injury （RILI） after involved-field intensity-modulated radiation therapy （IMRT） in patients with locally advanced non-small cell lung cancer （NSCLC）, and to evaluate the predictability of different models. Methods The clinical data of 242 inoperable or unresectable stage Ⅲ NSCLC patients treated in our hospital from 2007 to 2011 were reviewed. Grade 2 and grade 3 RILI that occurred within 6 months after IMRT were selected as outcome events in this study. The principal component analysis （PCA） model, Lyman-Kutcher-Burman （LKB） model, and mean lung dose （MLD） model were each used to establish a predictive model of normal tissue complication probability （NTCP） for evaluating the dosimetric parameters of IMRT. Results Four principal components were used in the PCA model. The areas under the receiver operating characteristic curve （AUCs） of grade 2 and grade 3 RILI were 0.652 and 0.611, respectively. For the LKB model, the fitted parameters were m=0.46, n=1.35, and D50=23.59 Gy for grade 2 RILI, and m=0.36, n=0.27, and D50=72.67 Gy for grade 3 RILI. The AUCs of grade 2 and grade 3 RILI in the LKB model were 0.607 and 0.585, respectively. For the MLD model, the estimated parameters were γ50=1.073 and D50=24.66 Gy for grade 2 RILI, and γ50=0.97 and D50=48.45 Gy for grade 3 RILI. The AUCs of grade 2 and grade 3 RILI in the MLD model were 0.604 and 0.569, respectively. Conclusions The use of large data set from a single patient population with the same mode of treatment is very important for improving model predictability and stability. Both the LKB model and PCA model can predict the probability of RILI, whereas the MLD model is less effective in predicting grade 3 RILI.