CT影像组学对结直肠癌肝转移的诊断价值

The diagnostic value of CT-based radiomics in liver metastasis of colorectal cancer

目的:探讨基于结直肠癌CT图像的影像组学方法在结直肠癌肝转移的诊断价值。方法:回顾性分析2017年6—12月吉林大学第一医院经手术病理证实为结直肠癌,并具有完整的术前结肠CT影像资料及术前常规检查等临床资料齐全的100例患者(肝转移、非肝转移患者各50例),按照4∶1的比例由计算机软件随机将患者分为建模组80例,测试组20例。采用Matlab 2017a软件和Python软件的特征提取算法提取影像组学特征,进行特征筛选并建立影像组学标签。通过纳入影像组学标签及患者的临床资料建立多变量随机森林分类器模型(Random forest classifier,RFC),并进行模型优化及验证,采用留出法、交叉验证法评价模型效能。结果:肝转移组与非肝转移组患者影像组学标签有显著差异(P<0.05)。影像组学标签、癌胚抗原(CEA)、糖类抗原19-9(CA19-9)表达水平均与结直肠癌肝转移诊断成正相关(P均<0.05)。RFC将建模组中训练集与验证集比例界定为7∶3,模型准确率为81.5%,训练集中ROC下面积(AUC)为0.991,敏感度为84.0%,特异度为96.8%,阳性预测值0.955,阴性预测值0.882;验证集中AUC为0.811,敏感度为72.7%,特异度为92.3%,阳性预测值为0.889,阴性预测值为0.800;RFC在十折交叉验证中准确率为81.0%。RFC对于测试组准确率为75.0%。结论:通过基于结直肠癌CT图像获得的影像组学标签及临床资料(CEA、CA19-9表达水平)建立的RFC,对于术前诊断结直肠癌肝转移有应用价值。

Objective： To explore the predictive value of a CT-based radiomics for liver metastasis in colorectal cancer. Methods： In the retrospective study, 100 patients with pathologically confirmed by surgery to colorectal cancer and preopera- tive contrast-enhanced CT examination in the First Hospital of Jilin University from June to December 2017 were included（50 cases with liver metastasis; 50 cases without liver metastasis）. The patients were divided into modeling group （80 cases） and testing group （20 cases） by computer random software according to the ratio of 4：1. Using the software matlab 2017a and python to extract a list of radiomies features and construct the corresponding radiomics signature. The radiomics signature and clinical variable were included to establish multivariable random forest classifier（RFC） model that was simplified and validated. The efficiency of the model was evaluated by the method of hold-out and cross validation. Results： The discrimination of ra- diomics signature between the liver metastasis group and non-liver metastasis group is significant （/9〈0.05）. The radiomics sig- nature, carcinoembryonic antigen （CEA）, carbohydrate antigen 19-9 （CA19-9） expression were showed positive correlation with the liver metastasis of colorectal cancer（P〈0.05）. The RFC splits modeling group into training dataset and validation dataset as 7：3, the accuracy is 81.5%, training dataset （AUC=0.991, sensitivity=84.0%, specificity=96.8%, positive predicted value=0.955, negative predictive value=0.882） and validation dataset（AUC=0.811, sensitivity=72.7%, specificity=92.3%, positive predicted val- ue=0.889, negative predictive value=0.800）. The accuracy in ten folded cross validation is 81.0%. The accuracy of testing group is 75.0%. Conclusion： The RFC model integrated with the radiomics signature based on CT imaging and clinical char- acteristic can be useful for diagnosis of liver metastasis in colorectal cancer.