CT影像组学模型鉴别诊断肺淋巴瘤与肺浸润性黏液腺癌

目的观察CT影像组学模型鉴别诊断肺淋巴瘤与肺浸润性黏液腺癌(PIMA)的价值。方法回顾性分析经病理证实的34例肺淋巴瘤(淋巴瘤组)及64例PIMA患者(PIMA组),按7∶3比例将其随机分入训练集和验证集。于胸部CT纵隔窗图像上沿病灶边缘手动勾画三维容积感兴趣区(VOI),提取影像组学特征参数。以最小冗余最大相关、LASSO十折交叉验证进行特征降维,以多因素Logistic回归分别构建影像组学标签、影像学特征模型及二者融合的个体化预测模型(以列线图表示)。采用受试者工作特征(ROC)曲线评价各模型对肺淋巴瘤与PIMA的鉴别效能,以决策曲线分析(DCA)综合评价模型的临床效用价值。结果共获得15个影像组学参数用于构建影像组学标签;ROC曲线结果显示其鉴别训练集肺淋巴瘤与PIMA的AUC=0.84,验证集AUC=0.77。以空气支气管征、支气管扩张及胸腔积液构建影像学特征模型,训练集AUC=0.85,验证集AUC=0.81;融合列线图对训练集AUC=0.95,验证集AUC=0.92;列线图具有更高的临床效用价值。结论基于CT征象、影像组学标签构建的个体化预测模型可有效鉴别肺淋巴瘤与PIMA。

CT radiomics models for differential diagnosis of pulmonary lymphoma and pulmonary invasive mucinous adenocarcinoma

Objective To investigate the value of CT radiomics models for differentiating pulmonary lymphoma and pulmonary invasive mucinous adenocarcinoma (PIMA). Methods Data of 34 patients with pulmonary lymphoma (lymphoma group) and 64 patients with PIMA (PIMA group) confirmed by pathology were retrospectively analyzed. The patients were randomly divided into training set and validation set according to 7:3 ratio. Volume of interest (VOI) was drawn through manually tracing the tumors on mediastinal window images of plain chest CT, and the radiomics features were extracted. The minimum redundancy and maximum correlation (mRMR), also Lasso 10-fold cross verification method were used to reduce the dimension of radiomics features. Multiple Logistic regression analysis was used to develop of radiomics signatures model and imaging features model, also individual prediction model (nomogram) through combining both the formers, respectively. Receiver operating characteristic (ROC) curve was used to evaluate the differential efficacy of 3 models for lung lymphoma and PIMA. The clinical utility value of models were evaluated with decision curve analysis (DCA). Results After dimensionality reduction, 15 radiomics feature parameters were obtained to construct radiomics signatures. In the training set, AUC of radiomics signatures for identifying lung lymphoma and PIMA was 0.84, while AUC in the validation set was 0.77. The air bronchial sign, bronchiectasis and pleural effusion on plain CT were incorporated into the imaging feature model, with AUC 0.85 in training set and 0.81 in validation set. AUC of nomogram in training set was 0.95, in validation set was 0.92. Nomogram had higher clinical value than the other models. Conclusion Individual prediction model based on plain CT features and radiomics signatures could effectively differentiate pulmonary lymphoma and PIMA.