基于剂量组学预测肺癌患者放射性肺炎发生的研究

目的 探讨剂量组学在预测肺癌根治性放疗患者放射性肺炎发生中的应用潜能。方法 回顾性收集行根治性放疗的314例肺癌患者的临床资料、放疗剂量文件、定位及随访CT图像,根据临床资料及影像学随访资料对放射性肺炎进行分级,提取全肺的剂量组学特征,构建机器学习模型。应用1000次自助抽样法（bootstrap）的最小绝对值收敛和选择算子嵌套逻辑回归（LASSO‐LR）及1000次bootstrap的赤池信息量准则（AIC）向后法筛选与放射性肺炎相关的剂量组学特征,随机按照7∶3划分为训练集及验证集,应用逻辑回归建立预测模型,并应用ROC曲线及校正曲线评价模型的性能。结果 共提取120个剂量组学特征,经LASSO‐LR降维筛选得到12个特征进入“特征池”,再经过AIC向后法筛选,最终筛选出6个剂量组学特征进行模型构建,训练集AUC为0.77（95%CI为0.65~0.87）,独立验证集AUC为0.72（95%CI为0.64~0.81）。结论 利用剂量组学建立的预测模型具有预测放射性肺炎发生的潜力,但仍需继续纳入多中心数据及前瞻性数据进一步挖掘剂量组学的应用潜能。

Dosiomics‐based prediction of incidence of radiation pneumonitis in lung cancer patients

Objective To explore the potential of dosiomics in predicting the incidence of radiation pneumonitis by extracting dosiomic features of definitive radiotherapy for lung cancer, and building a machine learning model. Methods The clinical data, dose files of radiotherapy, planning CT and follow‐up CT of 314 patients with lung cancer undergoing definitive radiotherapy were collected retrospectively. According to the clinical data and follow‐up CT, the radiation pneumonia was graded, and the dosiomic features of the whole lung were extracted to establish a machine learning model. Dosiomic features associated with radiation pneumonia by LASSO‐LR with 1000 bootstrap and AIC backward method with 1000 bootstraps were selected. Training cohort and validation cohort were randomly divided on the basis of 7:3.Logistic regression was used to establish the prediction model, and ROC curve and calibration curve were adopted to evaluate the performance of the model. Results A total of 120 dosiomic features were extracted. After LASSO‐LR dimensionality reduction, 12 features were selected into the "feature pool".After AIC, 6 dosiomic features were finally selected for model construction. The AUC of training cohort was 0.77(95%CI: 0.65 to 0.87), and the AUC of validation cohort was 0.72 (95%CI: 0.64 to 0.81). Conclusion The dosiomics prediction model has the potential to predict the incidence of radiation pneumonia, but it still needs to include multicenter data and prospective data.