Prediction of Aortic Contrast Enhancement on Dynamic Hepatic Computed Tomography - Performance Comparison of Machine Learning Methods and Simulation Software
Prediction of Aortic Contrast Enhancement on Dynamic Hepatic Computed Tomography - Performance Comparison of Machine Learning Methods and Simulation Software
复制标题
动态肝脏计算机断层扫描的主动脉对比度增强预测 - 机器学习方法和仿真软件的性能比较
DOI：
10.1097/rct.0000000000001273
复制
发表时间：
2022
影响因子：
1.3
通讯作者：
Awai Kazuo
中科院分区：
文献类型：
--
作者：
Masuda Takanori;Nakaura Takeshi;Higaki Toru;Funama Yoshinori;Sato Tomoyasu;Masuda Shouko;Yoshiura Takayuki;Arao Shinichi;Hiratsuka Junichi;Hirai Toshinori;Awai Kazuo
本研究的目的是比较集成机器学习(ML)方法和仿真软件对动态肝脏CT主动脉增强的预测能力。方法将339例人体肝脏动态CT扫描分为2组。其中一组由279次扫描组成，用于创建交叉验证数据集，另一组由60次扫描组成，用作测试数据集。为了评估患者特征对增强的影响，我们计算了肝动脉相腹主动脉每次增强时造影剂剂量的变化。ML的参数包括患者性别、年龄、身高、体重、体重指数和心输出量。我们通过5次交叉验证训练了9个ML回归变量，将所有ML回归变量的预测集成到集成学习和模拟中，并使用训练和测试数据比较它们的Pearson相关系数。结果不同ML方法的比较显示，每次腹主动脉增强的实际对比剂剂量和预测对比剂剂量的Pearson相关系数与集成ML最高(r=0.786)。这比用模拟软件得到的结果要高(r=0.350)。集合最大似然法的Bland-Altman一致性限[平均差值为5.26Hounsfield单位(HU)；95%一致性限为−112.88~123.40 HU]小于模拟软件的一致性限(平均值差值为11.70HU；95%一致性限为−164.71~188.11 HU)。结论集合最大似然法对肝动脉相腹主动脉增强的预测效果优于仿真软件。
ObjectivesThe aim of this study was to compare prediction ability between ensemble machine learning (ML) methods and simulation software for aortic contrast enhancement on dynamic hepatic computed tomography.MethodsWe divided 339 human hepatic dynamic computed tomography scans into 2 groups. One group consisted of 279 scans used to create cross-validation data sets, the other group of 60 scans were used as test data sets. To evaluate the effect of the patient characteristics on enhancement, we calculated changes in the contrast medium dose per enhancement of the abdominal aorta in the hepatic arterial phase. The parameters for ML were the patient sex, age, height, body weight, body mass index, and cardiac output. We trained 9 ML regressors by applying 5-fold cross-validation, integrated the predictions of all ML regressors for ensemble learning and the simulations, and used the training and test data to compare their Pearson correlation coefficients.ResultsComparison of different ML methods showed that the Pearson correlation coefficient for the real and predicted contrast medium dose per enhancement of the abdominal aorta was highest with ensemble ML (r= 0.786). It was higher than that obtained with the simulation software (r= 0.350). With ensemble ML, the Bland-Altman limit of agreement [mean difference, 5.26 Hounsfield units (HU); 95% limit of agreement,− 112.88 to 123.40 HU] was narrower than that obtained with the simulation software (mean difference, 11.70 HU; 95% limit of agreement,− 164.71 to 188.11 HU).ConclusionThe performance for predicting contrast enhancement of the abdominal aorta in the hepatic arterial phase was higher with ensemble ML than with the simulation software.