基于非增强MRI的影像组学术前预测肝细胞癌微血管浸润的研究

Predicting microvascular invasion of hepatocellular carcinoma based on non-contrast-enhanced MRI radiomics signature

目的建立基于非增强MRI的影像组学模型对肝细胞癌(hepatocellular carcinoma,HCC)的微血管侵犯(microvascularinvasion,MVI)进行术前预测。材料与方法回顾性分析经手术病理证实是否有MVI的HCC病人129例。所有患者术前2周内行3.0 T MRI。在T2WI-FS及ADC图中逐层勾画病灶区域提取影像组学特征。使用三步降维方法Variance Threshold、SelectKBest、LASSO算法依次进行降维来进行特征选择。分别使用六种分类器包括逻辑回归(logistic regression,LR)、支持向量机(supportvectormachine,SVM)、K近邻(K-NearestNeighbor,KNN)、决策树(decision tree,DT)、随机森林(random forest,RF)、极限梯度增强树(extreme gradient boosting,XGBoost)对提取的特征进行机器学习。通过绘制ROC曲线下面积(areaundercurve,AUC)、敏感度(sensitivity)、特异度(specificity)三个指标来评价各分类器所构建模型的效能。结果从T2WI-Fat suppressed (FS)及ADC图中分别提取出1409个影像组学特征。经过降维,最终从T2WI-FS图中筛选出12个以及从ADC图中选出8个最优特征来分别构建两个组学模型。两种分类器SVM、LR基于T2WI-FS特征所构建的模型性能最佳,对应的受试者工作特征AUC值分别为0.869、0.801,准确度为0.78、0.81。结论使用T2WI-FS的12个组学特征,可以获得较高的AUC值和准确度。因此,认为基于T2WI-FS的三维成像组学特征可以作为潜在的生物标志物来对肝细胞癌的微血管浸润进行术前非侵入性预测。

Objective: Magnetic resonance imaging(MRI)-based radiomics signatures was conducted to predict microvascular invasion(MVI) of hepatocellular carcinoma(HCC) preoperatively. Materials and Methods: One hundred and twenty-nine HCC patients who had undergone MRI examination on 3.0 T MRI were recruited. Radiomics features were extracted from fat-suppressed T2-weighted(T2 WI-FS) imaging and apparent diffusion coefficient(ADC) map. We used the Variance Threshold, SelectKBest, and least absolute shrinkage and selection operator(LASSO) algorithms in order to perform dimensionality reduction. Then random forests(RF), k-nearest neighbor(KNN), extreme gradient boosting(XGBoost), logistic regression(LR), decision tree(DT) and support vector machine(SVM) algorithm were trained to separate the HCC with MVI positive and with MVI negative. The performance of each model built by the classifier was evaluated by AUC and accuracy. Results: Quantitative imaging features(n=1409) were extracted from T2 WI-FS and ADC map respectively. Finally, 12 features of T2 WI-FS and 8 features of ADC were selected to construct the radiomics model separately. The model that used SVM classification method achieved the best performance among the six methods, with AUC values of 0.87, accuracy of 0.78 based on T2 WI-FS, and AUC values of 0.75, accuracy of 0.71 based on ADC. Conclusions: Good accuracy and AUC could be obtained using only 12 radiomic features of T2 WI-FS. Therefore, we proposed radiomics features from T2 WI-FS could be used as candidate biomarkers for preoperative prediction of MVI of HCC noninvasively.