基于机器学习的影像组学模型在重度无症状性颈动脉狭窄危险分层中的应用

Machine learning-based radiomics model for risk stratification of severe asymptomatic carotid stenosis

目的探讨基于机器学习的影像组学模型在重度无症状性颈动脉狭窄危险分层中的应用价值。方法回顾性收集2017—2021年中日友好医院心脏血管外科188例重度颈动脉狭窄患者的病例资料及头颈CT血管造影图像,其中训练集131例[男107例、女24例,平均年龄(68±8)岁],验证集57例[男50例、女7例,平均年龄(67±8)岁]。在横断面上沿颈动脉斑块的边缘逐层手动勾画感兴趣体积。使用Python软件的Pyradiomics包提取影像组学特征。采用组内及组间相关系数、冗余性分析、最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)回归分析进行特征筛选。使用逻辑回归、决策树、随机森林、支持向量机、朴素贝叶斯和K最邻近6种不同的有监督机器学习算法将筛选出的影像组学特征用于构建预测模型。采用受试者工作特征(receiver operating characteristic,ROC)曲线和ROC曲线下面积(area under the curve,AUC)比较各预测模型的诊断效能,并通过验证集进行验证。使用校准曲线和决策曲线分析(decision curve analysis,DCA)评价预测模型的校准度和临床实用性。结果基于训练集最终筛选出4个影像组学特征用于构建预测模型。在6种机器学习模型中,逻辑回归模型表现出较高且稳定的诊断效能,在训练集中的AUC为0.872,灵敏度为100.0%,特异性为66.2%;在验证集中的AUC为0.867,灵敏度为83.3%,特异性为78.8%。校准曲线及DCA显示,逻辑回归模型具有良好的校准度及临床应用价值。结论基于机器学习的影像组学预测模型在重度无症状性颈动脉狭窄患者危险分层中具有一定的应用价值。

Objective To explore the utility of machine learning-based radiomics models for risk stratification of severe asymptomatic carotid stenosis(ACS).Methods The clinical data and head and neck CT angiography images of 188 patients with severe carotid artery stenosis at the Department of Cardiovascular Surgery,China-Japan Friendship Hospital from 2017 to 2021 were retrospectively collected.The patients were randomly divided into a training set(n=131,including 107 males and 24 females aged 68±8 years),and a validation set(n=57,including 50 males and 7 females aged 67±8 years).The volume of interest was manually outlined layer by layer along the edge of the carotid plaque on crosssection.Radiomics features were extracted using the Pyradiomics package of Python software.Intraclass and interclass correlation coefficient analysis,redundancy analysis,and least absolute shrinkage and selection operator regression analysis were used for feature selection.The selected radiomics features were constructed into a predictive model using 6 different supervised machine learning algorithms:logistic regression,decision tree,random forest,support vector machine,naive Bayes,and K nearest neighbor.The diagnostic efficacy of each prediction model was compared using the receiver operating characteristic(ROC)curve and the area under the curve(AUC),which were validated in the validation set.Calibration and clinical usefulness of the prediction model were evaluated using calibration curve and decision curve analysis(DCA).Results Four radiomics features were finally selected based on the training set for the construction of a predictive model.Among the 6 machine learning models,the logistic regression model exhibited higher and more stable diagnostic efficacy,with an AUC of 0.872,a sensitivity of 100.0%,and a specificity of 66.2%in the training set;the AUC,sensitivity and specificity in the validation set were 0.867,83.3%and 78.8%,respectively.The calibration curve and DCA showed that the logistic regression model had good calibration and clinical usefulness.Conclusion The machine learning-based radiomics model shows application value in the risk stratification of patients with severe ACS.