基于放射组学预测放射性肺炎的初步研究

目的基于肺癌患者胸部定位CT图像进行筛选与放射性肺炎发生相关的放射组学特征,构建机器学习模型,探讨放射组学在预测放射性肺炎发生中的价值。方法回顾性分析行根治性调强放疗的Ⅲ期非小细胞肺癌患者86例,通过随访影像学资料及临床信息将其进行放射性肺炎分级,并收集其定位CT图像。将全肺作为感兴趣体积进行放射组学特征的提取,分析与发生放射性肺炎有关的放射组学特征及临床、剂量学特征。利用支持向量机进行模型构建,通过五折验证方式检测模型预测性能。结果提取出放射组学特征1029个,通过方差分析及LASSO方法共得到与发生放射性肺炎有关的放射组学特征5个。单纯利用放射组学特征构建模型的测试集曲线下面积(AUC)=0.67,利用放射组学特征结合临床、剂量参数构建模型的测试集AUC=0.71。结论在通过利用Ⅲ期非小细胞肺癌患者的定位CT图像进行构建的放射组学模型有预测放射性肺炎发生的潜能,加入临床及剂量参数后可进一步提高预测效能。

Preliminary study of predicting radiation pneumonitis based on radiomics technology

Objective To identify the radiomics features related to the occurrence of radiation pneumonitis based on localized CT images of the chest in lung cancer patients, establish a machine learning model and investigate the value of radiomics technology in predicting the incidence of radiation pneumonitis. Methods Clinical data of 86 patients with stage Ⅲ non-small cell lung cancer who received radical intensity-modulated radiation therapy (IMRT) were retrospectively analyzed. The radiation pneumonitis was graded by follow-up imaging data and clinical information. The planning CT images were collected. The lung was used as the volume of interest for extraction of radiomics features. The radiomics features,clinical and dosimetric parameters associated with the incidence of radiation pneumonitis were analyzed. Using the support vector machine to construct the model,the prediction performance of the model was evaluated by the five-fold verification method. Results A total of 1029 radiomics features were extracted from CT images and 5 features were selected by ANOVA and LASSO. Two validation sets showed differences between adopting radiomics features alone and incorporating clinical and dosimetric parameters and radiomics features (AUC=0.67 and 0.71,respectively). Conclusions The radiomics model constructed by planning CT images of lung cancer patients has the potential to predict the occurrence of radiation pneumonitis. Addition of clinical and dosimetric parameters can further improve the prediction performance of the model.