基于CT增强影像组学的支持向量机模型术前预测肝细胞癌微血管浸润的研究

Contrast-Enhanced CT Radiomics Combined with Support Vector Machine for Preoperatively Predicting Microvascular Invasion in Hepatocellular Carcinoma

目的探讨基于支持向量机模型(SVM)的CT增强影像组学方法对肝细胞癌(HCC)微血管浸润(MVI)的术前预测价值。方法回顾性分析186例经手术病理证实为HCC患者的临床及CT增强图像资料,其中MVI阳性83例,MVI阴性103例。首先对临床资料及影像学特征进行单因素及多因素分析,得到HCC发生MVI的独立危险因素。另外采用达尔文科研平台在CT增强动脉期、门静脉期及平衡期图像上进行影像组学特征提取及筛选。按照7∶3的比例将数据分为训练组和测试组,对训练组的组学特征构建SVM模型,并对MVI的独立预测因子构建Logistic回归模型。采用受试者工作特征(ROC)曲线分析模型的有效性,并用测试组进一步验证。采用Delong检验比较测试组中临床-影像学特征联合模型及不同期相影像组学模型的诊断效能。结果多因素Logistic回归分析得出3个MVI的独立预测因素:肿瘤最大径、肿瘤边缘不光滑及瘤内动脉。临床-影像学特征联合模型在训练组的诊断效能ROC曲线下面积(AUC)为0.884,测试组AUC值为0.753。CT增强影像组学特征筛选后分别在动脉期、门静脉期、平衡期及三期联合获得2、1、1及2个参数,包括二维最大直径(冠状位)和依赖熵,这两个特征在MVI阳性组和MVI阴性组有统计学差异(P<0.05)。采用SVM方法建立多期影像组学模型,动脉期、门静脉期、平衡期及增强三期联合模型在训练组的诊断效能AUC值分别为0.932、0.930、0.924及0.933,测试组AUC值分别为0.865、0.834、0.855及0.858。经Delong检验分析后,发现测试组中CT增强各期影像组学模型的诊断效能均优于临床-影像学特征联合模型,且动脉期组学模型的诊断效能较高。结论基于CT增强影像组学特征的SVM模型能够在术前无创地评估和预测MVI,可作为指导临床后续个性化治疗的有效工具。

Objective To investigate the preoperative predictive value of radiomics features combine with support vector machine(SVM) method for microvascular invasion(MVI) of hepatocellular carcinoma(HCC) based on contrast-enhanced CT imaging. Methods The clinical and contrast-enhanced CT image data of 186 cases with surgically pathologically confirmed HCC were retrospectively analyzed,including 83 cases with positive MVI and 103 cases with negative MVI. Firstly,the clinical data and imaging features were analyzed by univariate and multivariate analysis to obtain independent risk factors of MVI. In addition,the Darwin scientific research platform was used to extract and select radiomics features on CT arterial phase,portal venous phase and equilibrium phase images. All datas were divided into training group and test group according to the ratio of 7∶ 3. And support vector machine model was constructed for the radiomics features obtained from the training group. A Logistic regression model was build for the independent predictors of MVI. The effectiveness of the model was analyzed using the receiver operating characteristic curve(ROC),and its diagnostic efficacy was further verified using the test group,and the sensitivity,specificity and accuracy were calculated. The Delong test was used to compare the diagnostic efficacy of the combined clinical-imaging characteristics model and the different simultaneous radiomics models in test group. Results Multivariate Logistic regression analysis showed that maximum tumor diameter,non-smooth tumor margin and intratumoral artery were independent predictors of MVI. The AUC of the combined clinical-imaging features model in the training group and test group was 0. 884 and 0. 753,respectively. There were 2,1,1 and 2 parameters which selected from contrast-enhanced CT radiomics features in the arterial phase,portal phase,equilibrium phase and three phases combined,respectively,including two-dimensional maximum diameter(column) and dependence entropy,which were statistically different in the MVI-positive and MVI-negative groups(P < 0. 05). The radiomics model was developed using SVM method. And the AUCs of arterial,portal venous,equilibrium and three-phase combined models were 0. 932,0. 930,0. 924 and 0. 933 in the training group and 0. 865,0. 834,0. 855 and 0. 858 in the test group,respectively. After analyzing by Delong validation,it was found that the diagnostic efficacy of contrast-enhanced CT radiomics model in each phase was better than the combined clinical-imaging feature model in the test group,and the diagnostic efficacy of the arterial phase radiomics model was higher. Conclusion Combining support vector machine approach with contrast-enhanced CT radiomics features could preoperative predict the occurrence of MVI in HCC patients noninvasively and can be used as an effective tool to guide the subsequent clinical personalized treatment.