基于CT的影像组学结合腹部脂肪面积预测肝细胞癌微血管侵犯

Prediction of Microvascular Invasion in Hepatocellular Carcinoma Using CT-Based Radiomics Combined with the Abdominal Fat Area

目的 探索基于CT门静脉期的影像组学模型结合腹部脂肪面积参数构建的联合模型对肝细胞癌(HCC)微血管侵犯(MVI)的预测价值。方法 回顾性搜集134例经病理证实是否存在MVI的HCC患者的平扫及增强CT图像。根据随机分层抽样的原则将患者以7∶3的比例划分为训练集(n=93)和测试集(n=41),在CT平扫图像上测量所有患者的腹部脂肪面积,通过单因素及多因素逻辑回归构建脂肪模型。通过自动化特征提取算法提取门静脉期肿瘤的影像组学特征,采用Spearman相关性分析和最小绝对收缩和选择算子(LASSO)进行特征筛选,并分别构建六种机器学习模型,将性能较优模型联合腹部脂肪面积参数构建联合模型。采用受试者工作特征曲线(ROC)曲线下面积(AUC)来评估模型的预测效能,校准曲线来验证校准能力,决策曲线用于分析和比较模型的临床实用性,Delong检验评估各模型之间预测效能的差异。结果 单因素及多因素逻辑回归提示内脏及皮下脂肪面积参数均为HCC-MVI的独立危险因素,以此构建的脂肪模型在训练集和测试集的AUC值为0.747(0.648~0.845)和0.696(0.536~0.857)。基于门静脉期肿瘤提取的影像组学特征经过逐层特征筛选及降维方法后得到10个影像组学特征用于六种机器学习模型的构建,其中支持向量机(SVM)模型的预测性能较好,在训练集和测试集中的AUC值分别为0.904(0.844~0.965)和0.838 (0.717~0.959),与腹部脂肪参数联合后构建联合模型,并用列线图对其进行可视化。结果显示,联合模型在训练集中的性能0.925(0.872~0.978)显著高于影像组学模型(P=0.029)和脂肪模型(P=0.0054),测试集中该模型性能0.877(0.772~0.983)显著高于脂肪模型(P=0.0165),高于影像组学模型(P=0.207),校准曲线显示模型拟合良好,决策曲线提示联合模型较其他模型具有较好的临床实用价值。结论 基于CT门静脉期的影像组学特征联合腹部脂肪面积构建的联合模型对HCC-MVI具有较高的预测价值。

Objective To explore the predictive value of a combined radiomics model and abdominal fat area parameters for microvascular invasion of hepatocellular carcinoma based on CT portal phase. Methods Retrospectively collected 134 cases confirmed by pathology exists Microvascular invasion of Hepatocellular carcinoma scan and enhanced CT images of patients. According to the principle of random stratified sampling, the patients were divided into a training set(n=93) and a testing set(n=41) with a ratio of 7∶3.Firstly, the abdominal fat area parameters of all patients were measured on CT images, and the fat model was constructed by univariate and multivariate Logistic regression. Then, radiomics features of portal venous tumors were extracted by an automated feature extraction algorithm. Spearman correlation analysis and the Least absolute shrinkage and selection operator were used for feature selection. Six kinds of machine learning models were constructed respectively. The better performance model combined with abdominal fat area parameters was used to construct the final combined model. The Area under the curve of the Receiver operating characteristic curve was used to evaluate the predictive efficacy of the model, and the calibration curve was used to verify the calibration capability. The Decision curve analysis is used to analyze and compare the clinical utility of models, and the Delong test assesses the differences in predictive power between models. Results Univariate and multivariate logistic regression indicated that Subcutaneous adipose tissue area and Visceral adipose tissue area parameters were independent risk factors for HCC-MVI,and the AUC values of the fat model constructed by this method were 0.747(0.648-0.845) and 0.696(0.536-0.857) in the training and test sets. After a series of feature screening and dimension reduction methods, The remaining 10 radiomics features were used for the construction of six machine learning models, the support vector machine model had better performance. In the training set and the testing set, the AUC values were 0.904(0.844-0.965) and 0.838(0.717-0.959),respectively. The joint model was constructed by combining the fat parameters and the radiomics model and visualized by using the nomogram. The performance of the combined model in the training set was 0.925(0.872-0.978),which was significantly higher than that of the radiomics model(P=0.029) and the fat model(P=0.0054).The performance of the combined model in the testing set was 0.877(0.772-0.983),which was significantly higher than that of the fat model(P=0.0165),which was higher than that of the radiomics model(P=0.207),the calibration curve showed that the model was well-fitted, and the analysis of the decision curve suggested that the joint model had better clinical practical value than other models. Conclusion The joint model based on radiomics features of the CT portal phase and the abdominal fat area has a high predictive value for HCC-MVI.