CT影像组学联合征象鉴别新型冠状病毒肺炎与其他病毒性肺炎的价值

The value of CT signs combined with radiomics in the differentiation of COVID-19 from other viral pneumonias

目的:探讨基于CT影像征象联合影像组学模型鉴别新型冠状病毒肺炎(COVID-19)和其他病毒性肺炎的临床价值。方法:回顾性分析2015年3月至2020年3月云南省15家医院实时逆转录聚合酶链反应检测为病毒性肺炎并接受胸部CT扫描的181例患者的临床和影像资料。根据患者的病毒类型分为COVID-19组(89例)和非COVID-19组(92例);所有病例按7∶3的比例随机分层抽样分为训练集(126例)和测试集(55例)。从首诊平扫胸部CT图像中提取出肺炎征象和影像组学特征,分别建立独立预测模型和联合预测模型。通过受试者操作特征(ROC)曲线、连续净重新分类指数(NRI)校准曲线和决策曲线分析各种模型的诊断性能。结果:联合模型由3个重要的CT征象和14个筛选的影像组学特征构成。影像组学模型在训练集中鉴别诊断COVID-19组和非COVID-19组的ROC曲线下面积(AUC)为0.904,灵敏度为85.5%,特异度为84.4%,准确度为84.9%;在测试集中鉴别诊断COVID-19组和非COVID-19组的AUC为0.866,灵敏度为77.8%,特异度为78.6%,准确度为78.2%。联合模型在训练集中鉴别诊断COVID-19组和非COVID-19组的AUC为0.956,灵敏为91.9%,特异度为85.9%,准确度为88.9%;在测试集中鉴别诊断COVID-19组和非COVID-19组的AUC为0.943,灵敏度为88.9%,特异度为85.7%,准确度为87.3%。联合模型与影像组学模型鉴别诊断COVID-19组与非COVID-19组的AUC值在训练集中差异具有统计学意义( Z=-2.43, P=0.015),在测试集中差异无统计学意义( Z=-1.73, P=0.083),进一步采用连续NRI分析,结果显示训练集和测试集中联合模型较单独影像组学模型均具有正向改善能力(训练集:连续NRI为1.077,95%CI 0.783~1.370;测试集:连续NRI为1.421,95%CI 1.051~1.790)。校准曲线表明,训练集和测试集中,联合模型预测COVID-19的概率与观察值之间具有良好的一致性;决策曲线显示联合模型阈值概率0~0.75时,可获得大于0.6的净收益。 结论:基于胸部CT影像征象联合影像组学模型鉴别诊断COVID-19和其他病毒性肺炎具有较好的临床价值。

Objective To explore the classification performance of combined model constructed from CT signs combined with radiomics for discriminating COVID-19 pneumonia and other viral pneumonia.Methods The clinical and CT imaging data of 181 patients with viral pneumonia confirmed by reverse transcription-polymerase chain reaction in 15 hospitals of Yunnan Province from March 2015 to March 2020 were analyzed retrospectively.The 181 patients were divided into COVID-19 group(89 cases)and non-COVID-19 group(92 cases),which were further divided into training cohort(126 cases)and test cohort(55 cases)at a ratio of 7∶3 using random stratified sampling.The CT signs of pneumonia were determined and the radiomics features were extracted from the initial unenhanced chest CT images to build independent and combined models for predicting COVID-19 pneumonia.The diagnostic performance of the models were evaluated using receiver operating characteristic(ROC)analysis,continuous net reclassification index(NRI)calibration curve and decision curve analysis.Results The combined models consisted of 3 significant CT signs and 14 selected radiomics features.For the radiomics model alone,the area under the ROC curve(AUC)were 0.904(sensitivity was 85.5%,specificity was 84.4%,accuracy was 84.9%)in the training cohort and 0.866(sensitivity was 77.8%,specificity was 78.6%,accuracy 78.2%)in the test cohort.After combining CT signs and radiomics features,AUC of the combined model for the training cohort was 0.956(sensitivity was 91.9%,specificity was 85.9%,accuracy was 88.9%),while that for the test cohort was 0.943(sensitivity was 88.9%,specificity was 85.7%,accuracy was 87.3%).The AUC values of the combined model and the radiomics model in the differentiation of COVID-19 group and the non-COVID-19 group were significantly different in the training cohort(Z=-2.43,P=0.015),but difference had no statistical significance in the test cohort(Z=-1.73,P=0.083),and further analysis using the NRI showed that the combined model in both the training cohort and the test cohort had a positive improvement ability compared with radiomics model alone(training cohort:continuous NRI 1.077,95%CI 0.783-1.370;test cohort:continuous NRI 1.421,95%CI 1.051-1.790).The calibration curve showed that the prediction probability of COVID-19 predicted by the combined model was in good agreement with the observed value in the training and test cohorts;the decision curve showed that a net benefit greater than 0.6 could be obtained when the threshold probability of the combined model was 0-0.75.Conclusion The combination of CT signs and radiomics might be a potential method for distinguishing COVID-19 and other viral pneumonia with good performance.