Department of Radiation and Medical Oncology, The 1st Affiliated Hospital of Wenzhou Medical University, No.2 Fuxue Lane, Wenzhou, 325000, People's Republic of China.
Department of Medical Engineering, The 1st Affiliated Hospital of Wenzhou Medical University, Shangcai Village, Wenzhou, 325000, People's Republic of China.
Eur Radiol. 2019 Nov;29(11):6080-6088. doi: 10.1007/s00330-019-06193-w. Epub 2019 Apr 26.
To investigate the treatment response prediction feasibility and accuracy of an integrated model combining computed tomography (CT) radiomic features and dosimetric parameters for patients with esophageal cancer (EC) who underwent concurrent chemoradiation (CRT) using machine learning.
The radiomic features and dosimetric parameters of 94 EC patients were extracted and modeled using Support Vector Classification (SVM) and Extreme Gradient Boosting algorithm (XGBoost). The 94-sample dataset was randomly divided into a 70-sample training subset and a 24-sample independent test set while keeping the class proportions intact via stratification. A receiver operating characteristic (ROC) curve was used to assess the performance of models using radiomic features alone and using combined radiomic features and dosimetric parameters.
A total of 42 radiomic features and 18 dosimetric parameters plus the patients' characteristic parameters were extracted for these 94 cases (58 responders and 36 non-responders). XGBoost plus principal component analysis (PCA) achieved an accuracy and area under the curve of 0.708 and 0.541, respectively, for models with radiomic features combined with dosimetric parameters, and 0.689 and 0.479, respectively, for radiomic features alone. Image features of GlobalMean X.333.1, Coarseness, Skewness, and GlobalStd contributed most to the model. The dosimetric parameters of gross tumor volume (GTV) homogeneity index (HI), Cord Dmax, Prescription dose, Heart-Dmean, and Heart-V50 also had a strong contribution to the model.
The model with radiomic features combined with dosimetric parameters is promising and outperforms that with radiomic features alone in predicting the treatment response of patients with EC who underwent CRT.
• The model with radiomic features combined with dosimetric parameters is promising in predicting the treatment response of patients with EC who underwent CRT. • The model with radiomic features combined with dosimetric parameters (prediction accuracy of 0.708 and AUC of 0.689) outperforms that with radiomic features alone (best prediction accuracy of 0.625 and AUC of 0.412). • The image features of GlobalMean X.333.1, Coarseness, Skewness, and GlobalStd contributed most to the treatment response prediction model. The dosimetric parameters of GTV HI, Cord Dmax, Prescription dose, Heart-Dmean, and Heart-V50 also had a strong contribution to the model.
使用机器学习研究联合计算机断层扫描（CT）放射组学特征和剂量学参数的综合模型对接受同期放化疗（CRT）的食管癌（EC）患者的治疗反应预测的可行性和准确性。
提取 94 例 EC 患者的放射组学特征和剂量学参数，并使用支持向量分类（SVM）和极端梯度提升算法（XGBoost）进行建模。通过分层保持类比例不变，将 94 个样本数据集随机分为 70 个样本训练子集和 24 个独立测试集。使用受试者工作特征（ROC）曲线评估仅使用放射组学特征和使用联合放射组学特征和剂量学参数的模型的性能。
对这 94 例患者（58 例为应答者和 36 例为非应答者）共提取了 42 个放射组学特征和 18 个剂量学参数以及患者特征参数。XGBoost 加主成分分析（PCA）在结合放射组学特征和剂量学参数的模型中，获得了 0.708 的准确性和 0.541 的曲线下面积，在仅使用放射组学特征的模型中，获得了 0.689 的准确性和 0.479 的曲线下面积。放射组学特征中 GlobalMean X.333.1、粗糙度、偏度和全局标准差对模型的贡献最大。剂量学参数的肿瘤总体积（GTV）均匀性指数（HI）、 Cord Dmax、处方剂量、心脏-Dmean 和心脏-V50 对模型也有很大的贡献。
与仅使用放射组学特征的模型相比，联合放射组学特征和剂量学参数的模型在预测接受 CRT 的 EC 患者的治疗反应方面具有潜力。
联合放射组学特征和剂量学参数的模型在预测接受 CRT 的 EC 患者的治疗反应方面具有潜力。
联合放射组学特征和剂量学参数的模型（预测准确性为 0.708，AUC 为 0.689）优于仅使用放射组学特征的模型（最佳预测准确性为 0.625，AUC 为 0.412）。
放射组学特征中 GlobalMean X.333.1、粗糙度、偏度和全局标准差对治疗反应预测模型的贡献最大。剂量学参数的 GTV HI、Cord Dmax、处方剂量、心脏-Dmean 和心脏-V50 对模型也有很大的贡献。
