基于平扫CT三维影像组学特征的机器学习模型检测主动脉夹层的价值

Value of machine learning model for detecting aortic dissection based on non-contrast CT 3D-radiomics feature

目的探讨基于平扫CT三维影像组学特征的机器学习模型检测主动脉夹层(aortic dissection,AD)的价值,并根据Stanford分型分析其检测Stanford A型AD的价值。方法回顾性分析2011年7月至2022年7月间在复旦大学附属华东医院同时接受胸腹平扫和增强CT检查的患者资料。放射科医师根据增强CT图像中的表现来诊断AD。共纳入患者128例,其中AD阳性61例,AD阴性67例。采用简单随机化法,按照7∶3的比例将患者划分为训练集(n=89)和验证集(n=39)。采用3D Slicer在平扫图像上对主动脉手动勾画感兴趣区域体积(volumetric region of interest,VOI),Pyradiomics提取影像组学特征。用Spearman相关系数和最小绝对收缩选择算子(least absolute shrinkage and selection operator,LASSO)算法选取最优特征集,在此基础上分别构建支持向量机(support vector machine,SVM)、决策树、随机森林、极端梯度提升(extreme gradient boosting,XGBoost)、轻量梯度提升(light gradient boosting,LightGBM)和极端随机树模型,在验证集中分别对其检测AD和Stanford A型AD的效能进行验证。评价指标包括受试者操作特征曲线下面积(area under the receiver operating characteristic curve,AUC)、F1分数、准确度、召回率。结果在训练集中筛选出35个影像组学特征,基于这35个组学特征建立的机器学习模型中,XGBoost模型在检测AD上性能最优;在验证集中,XGBoost模型的AUC、F1分数、准确度和召回率分别为0.982、0.960、96.2%和100.0%。进一步探究影像组学机器学习模型对Stanford A型AD的检测效能时,在训练集中筛选出14个影像组学特征,在基于这14个组学特征构建的机器学习模型中,随机森林模型的检测效能最好;在验证集中,随机森林模型的AUC、F1分数、准确度和召回率分别为0.852、0.625、76.9%和100.0%。结论基于平扫CT三维影像组学特征的机器学习模型可用于检测AD及Stanford A型AD。

Objective To investigate the value of non-contrast CT 3D-radiomics based machine learning model for detecting aortic dissection(AD),and Stanford type A AD.Methods A total of 128 patients who were suspected AD and underwent both thoracic and abdominal non-contrast and enhanced CT examination in Huadong Hospital,Fudan University between Jul 2011 and Jul 2022 were retrospectively enrolled.Radiologists made a diagnosis of AD based on the presentation of the enhanced CT images.The patients were randomly divided into the training set(n=89)and the validation set(n=39)in a 7∶3 ratio with simple random sampling.3D slicer was used to manually delineate volumetric region of interest(VOI)on the non-contrast CT images.Pyradiomics was utilized to extract radiomics features from these images.The optimal feature set was selected using Spearman correlation coefficient and the least absolute shrinkage and selection operator(LASSO)algorithm.Subsequently,support vector machine(SVM),decision tree,random forest,eXtreme gradient boosting(XGBoost),light gradient boosting(LightGBM),and extra trees models were individually constructed based on this optimal feature set.The evaluation metrics included the area under the receiver operating characteristic curve(AUC),F1 score,accuracy,and recall.Results In the training set,35 radiomics features were selected,and based on these 35 features,the XGBoost model demonstrated the best performance in detecting AD.In the validation set,the XGBoost model achieved an AUC of 0.982,an F1 score of 0.960,an accuracy of 96.2%,and a recall of 100.0%in AD detection.Further for detecting Stanford type A AD,the optimal feature set consisted of 14 radiomic features in the training set.In the validation set,random forest model had the best performance,with the AUC of 0.852,F1 score of 0.625,accuracy of 76.9%and recall of 100.0%.Conclusion The non-contrast CT 3D-radiomics based machine learning model is valuable in detecting AD and Stanford type A AD.