基于超声影像组学建立肝棘球蚴病分型模型的可行性研究

Feasibility of ultrasound radiomics-based models for classification of hepatic echinococcosis

目的 探究基于超声影像组学构建肝棘球蚴病分型模型的可行性,从而为肝棘球蚴病精准超声诊断提供参考依据。方法 回顾性收集2014年10月于四川省甘孜藏族自治州石渠县采集的200例肝棘球蚴病患者超声声像图,勾画肝棘球蚴病病灶感兴趣区域。采用25种方法提取肝棘球蚴病影像组学特征,应用预选方式与最小绝对收缩和选择算法进行特征筛选,按7∶3比例将图像根据病灶类型随机划分为训练集与独立测试集。基于内核逻辑回归(kernel logistic regression, KLR)与高斯核函数型支持向量机(medium Gaussian support vector machine, MGSVM)两种分类器构建肝棘球蚴病分型的机器学习模型,绘制受试者工作特征(receiver operating characteristic, ROC)曲线,计算构建的机器模型用于肝棘球蚴病分型的敏感度、特异度及曲线下面积(area under the curve, AUC)。结果 25种方法累计提取5 005个棘球蚴病患者超声影像组学特征,经特征选择筛选出36个最优影像组学特征,并在此基础上建立了KLR和MGSVM两种机器学习模型。ROC曲线分析显示,MGSVM模型在训练集中用于肝棘球蚴病分型效果更优,敏感度、特异度和AUC分别为0.82、0.78和0.88,而KLR模型在独立测试集中表现更佳,敏感度、特异度和AUC分别为0.82、0.72和0.86。结论 基于超声影像组学的机器学习模型可用于肝棘球蚴病分型。

ObjectiveTo investigate the feasibility of establishment of ultrasound radiomics-based models for classification of hepatic echinococcosis, so as to provide insights into precision ultrasound diagnosis of hepatic echinococcosis.MethodsThe ultrasonographic images were retrospectively collected from 200 patients with hepatic echinococcosis in Shiqu County, Ganzi Tibetan Autonomous Prefecture, Sichuan Province in October 2014, and the regions of interest were plotted in ultrasonographic images of hepatic echinococcosis lesions. The ultrasound radiomics features of hepatic echinococcosis were extracted with 25 methods, and screened using pre-selection and the least absolute shrinkage and selection operator. Then, all ultrasonographic images were randomly assigned into the training and independent test sets according to the type of lesions at a ratio of 7∶3. Machine learning models for classification of hepatic echinococcosis were created based on two classifiers, including kernel logistic regression(KLR) and medium Gaussian support vector machine(MGSVM). The receiver operating characteristic(ROC) curves were plotted, and the sensitivity, specificity and areas under the curves(AUC) of the created machine learning models for classification of hepatic echinococcosis were calculated.ResultsA total of 5 005 ultrasound radiomics features were extracted from 200 patients with hepatic echinococcosis using 25 methods, and 36 optimal radiomics features were screened through feature selection,based on which two machine learning models were created, including KLR and MGSVM. ROC curve analysis showed that MGSVM presented a higher efficacy for hepatic echinococcosis classification than KLR in the training set, with a sensitivity of 0.82, a specificity of 0.78 and AUC of 0.88, while KLR presented a higher efficacy for hepatic echinococcosis classification than MGSVM in the independent test set, with a sensitivity of 0.82, a specificity of 0.72 and AUC of 0.86, respectively.ConclusionUltrasound radiomics-based machine learning models are feasible for hepatic echinococcosis classification.