CT影像组学在特发性肺纤维化患者预后评估中价值

Value of CT-based radiomics to the prognostic evaluation of patients with idiopathic pulmonary fibrosis

目的基于CT影像组学构建特发性肺纤维化(IPF)预后预测模型,探讨其在IPF患者预后评估中的价值。方法2018年1月1日—2020年12月31日河南省人民医院诊治IPF患者171例,均采用吡非尼酮或尼达尼布抗纤维化治疗,随访2年,根据随访期间预后情况分为预后不良者101例和预后良好者70例。按照7∶3比例将171例患者随机分为训练集119例和验证集52例,比较训练集与验证集患者体质量指数及吸烟史、特殊物质接触史、预后不良比率。2组入院时均行胸部高分辨率CT检查,对CT影像图中蜂窝影进行感兴趣区标注,应用Onekeyai platform软件进行特征提取,采用lasso回归筛选影像特征并构建K最邻近法(KNN)、支持向量机(SVM)和LightGBM 3种模型;绘制ROC曲线,评估训练集和验证集中3种模型预测IPF患者预后不良的效能。结果训练集体质量指数及吸烟史、特殊物质接触史、预后不良比率与验证集比较差异均无统计学意义(P>0.05)。经特征提取共提取1648个影像特征,其中一阶特征342个,形状特征14个,纹理特征1292个(灰度共生矩阵418个、灰度行程矩阵304个、灰度区域大小矩阵304个和灰度差分矩阵266个);经特征筛选共筛选出76个影像特征;采用lasso回归最终纳入10个影像特征。训练集中SVM、KNN和LightGBM模型预测IPF患者预后不良的AUC分别为0.828(95%CI:0.754~0.903,P<0.001)、0.802(95%CI:0.728~0.877,P<0.001)、0.880(95%CI:0.815~0.944,P<0.001),灵敏度分别为87.1%、74.3%、71.4%,特异度分别为61.2%、73.5%、93.8%,准确率分别为75.6%、74.0%、79.0%;验证集中SVM、KNN和LightGBM模型预测IPF患者预后不良的AUC分别为0.897(95%CI:0.810~0.984,P=0.044)、0.685(95%CI:0.540~0.830,P=0.025)、0.867(95%CI:0.761~0.973,P<0.001),灵敏度分别为80.6%、100.0%、90.3%,特异度分别为85.7%、28.6%、81.0%,准确率分别为75.0%、61.5%、78.8%。结论基于CT影像组学构建的预测模型在IPF患者预后评估中有一定价值,SVM模型、LightGBM模型预测效能较高。

Objective To construct a prognostic prediction model of idiopathic pulmonary fibrosis(IPF)based on CT radiomics,and to explore its value to the prognosis evaluation of IPF.Methods From January 1,2018 to December 31,2020,171 patients with IPF received anti-fibrosis treatment with pirfenidone or nintedanib in Henan Provincial People's Hospital.According to the follow-up results two years later,171 patients were divided into poor prognosis group(n=101)and good prognosis group(n=70).Based on a ratio of 7:3,171 patients were randomly divided into the training set(n=119)and the validation set(n=52).The body mass index,history of smoking,history of exposure to special substances,and rate of poor prognosis were compared between two sets.Both two groups were performed high-resolution CT scan on admission.The region of interest was marked in the CT image.Onekeyai platform software was used for feature extraction,and lasso regression analysis was done to filter the image features and construct three models of the K-Nearest Neighbor(KNN),Support Vector Machine(SVM)and LightGBM.ROC curves were plotted to evaluate the efficiencies of these 3 models on predicting poor prognosis in two sets.Results There were no significant differences in the body mass index,history of smoking,history of exposure to special substances,and rate of poor prognosis between two sets(P>0.05).Totally 1648 image features were extracted by feature extraction,including 342first-order features,14 shape features,and 1292 texture features(418 gray-level co-occurrence matrices,304 gray-level run-length matrices,304 gray-level size zone matrices and 266 gray-level tone difference matrices);a total of 76 image features were selected by feature selection;10 image features were finally included by adopting lasso regression.The AUCs of SVM,KNN and LightGBM models in the training set for predicting poor prognosis of IPF patients were 0.828(95%CI:0.754-0.903,P<0.001),0.802(95%CI:0.728-0.877,P<0.001)and 0.880(95%CI:0.815-0.944,P<0.001),the sensitivities were 87.1%,74.3%and 71.4%,the specificities were 61.2%,73.5%and 93.8%,and the accuracy rates were 75.6%,74.0%and 79.0%,respectively.The AUCs of SVM,KNN and LightGBM models in the validation set were 0.897(95%CI:0.810-0.984,P=0.044),0.685(95%CI:0.540-0.830,P=0.025)and 0.867(95%CI:0.761-0.973,P<0.001),the sensitivities were 80.6%,100.0%and 90.3%,the specificities were 85.7%,28.6%and 81.0%,and the accuracy rates were 75.0%,61.5%and 78.8%,respectively.Conclusion The prediction model based on CT radiomics has a certain value in the prognosis evaluation of IPF patients,and SVM model and LightGBM model have relatively higher prediction efficiencies.