基于胸部CT图像和临床特征构建预测COVID-19并发心肌损伤的分类模型及其效能对比  

Construction and comparison of multiple classification models based on chest CT images and clinical cha- racteristics for predicting myocardial injury in hospitalized patients with COVID-19

作　　者：王荣华 王司琪 余卓 李瑞 武志峰 郭东强 WANG Rong-hua;WANG Si-qi;YU Zhuo(Department of Radiology,Shanxi Bethune Hospital,Shanxi Academy of Medical Sciences,Tongji Shanxi Hospital,the Third Hospital of Shanxi Medical University,Taiyuan 030032,China)

机构地区：[1]山西白求恩医院(山西医学科学院同济山西医院)/山西医科大学第三医院放射科,山西太原030032 [2]慧影医疗科技(北京)股份有限公司,北京100080

出　　处：《放射学实践》2024年第12期1585-1592,共8页Radiologic Practice

基　　金：山西白求恩医院/山西医学科学院“新冠感染应急攻关科研专项”基金(2023xg06-2)。

摘　　要：目的:基于胸部CT图像和临床资料构建并验证针对新型冠状病毒肺炎(COVID-19)患者并发心肌损伤的早期预测模型。方法:回顾性搜集2022年11月-2023年2月在本院行胸部CT平扫检查并经实验室确诊为COVID-19的382例住院患者的病例资料。定义高敏肌钙蛋白水平(Hs-cTnI)高于17.5 pg/mL为心肌损伤。其中,心肌损伤组143例,非心肌损伤组239例。分析两组患者的一般临床资料、实验室检查结果及胸部CT图像上四项定量参数值(包括升主动脉直径、肺动脉干直径、肺部炎症指数及肺部感染占比)的差异。入组患者以8∶2的比例随机纳入训练集(306例)和验证集(76例)。基于CT定量参数、临床指标和两者的组合,采用机器学习中的逻辑回归算法分别构建预测心肌损伤的CT模型、临床模型及CT-临床联合模型;此外,采用深度学习方法构建并验证基于CT图像的预测模型,输出热力图可视化深度学习模型关注的影像区域是否包含临床关注的异常影像区域,以评估模型的可解释性及合理性。在验证集中通过受试者工作特征曲线下面积(AUC)对所构建的各种模型的预测效能进行对比分析。结果:深度学习模型在预测COVID-19相关心肌损伤方面的效能最高,在验证集的AUC为0.970,高于CT放射学模型(AUC=0.735)、临床模型(AUC=0.907)和CT-临床联合模型(AUC=0.920)。热力图显示模型认为与心肌损伤分类相关的图像区域与CT图像上肺部异常表现区在视觉上具有良好的主观一致性。结论:基于胸部CT图像的深度学习模型为COVID-19相关心肌损伤的预测提供了一种及时、无创且较准确的手段,可为临床医师进行准确的患者分层和决策提供额外信息。Objection:To construct and validate classification models based on chest CT images and clinical characteristics for timely predicting myocardial injury in patients with COVID-19.Methods:We retrospectively collected 382 laboratory-confirmed hospitalized patients with COVID-19 from November 2022 to February 2023 who underwent chest CT.Myocardial injury was defined as high-sensitivity cardiac troponin I(Hs-cTnI)levels above 17.5pg/mL.There were 143 patients with myocardial injury and 239 patients without.Clinical date,laboratory results and four quantitative parameters on chest CT images(including ascending aortic diameter,main pulmonary artery diameter,pulmonary inflammation index,and pulmonary infection proportion)were analyzed between the two groups.All patients were randomly assigned to the training set(306 cases)and the validation set(76 cases)according to a ratio of 8∶2.Logistic regression algorithm in machine learning was employed to construct CT model,clinical model and CT-clinical combined model respectively based on quantitative CT parameters,clinical data and their combination for predicting myocardial injury.Additionally,a deep learning(DL)model based on CT images was developed and validated,with heat maps visualizing the regions of interest identified by the DL model to assess the model's interpretability and ratio-nality.The predictive efficacy of the three models was compared based on the area under the receiver operation characteristic curve(AUC).Results:The DL model showed good discrimination performance in the validation cohort(AUC=0.970),which was higher than the CT model(AUC=0.735),clinical model(AUC=0.907)and CT-clinical model(AUC=0.920).The heat map showed a good subjective consistency between the image areas considered relevant by the DL model for myocardial injury classification and those areas of interest in clinical diagnosis.Conclusion:The deep learning model based on chest CT images provides a timely,non-invasive,and relatively accurate means for predicting COVID-19 associated myocardia

关 键 词：新型冠状病毒肺炎 心肌损伤 深度学习 体层摄影术 X线计算机 影像组学 预测模型 

分 类 号：R814.42[医药卫生—影像医学与核医学] R542.2[医药卫生—放射医学] R563.1[医药卫生—临床医学]