基于3D CNN的人工智能辅助诊断模型对超声心动图视频异常改变的识别

Recognition of abnormal changes in echocardiographic videos by an artificial intelligence assisted diagnosis model based on 3D CNN

目的:探讨基于三维卷积神经网络(3D CNN)构建的人工智能辅助诊断模型对高血压性心脏病、出现心脏改变的慢性肾功能衰竭(CRF)及甲状腺功能减退症(甲减)患者超声心动图视频的诊断效能及其临床应用价值。方法:本研究为回顾性研究。收集2019年4月至2021年10月就诊于河南省人民医院的高血压性心脏病、出现心脏改变的CRF和甲减患者。依据诊断分为高血压组、CRF组和甲减组,另采用简单随机抽样方法从同期健康体检者中入选正常对照组。收集入选患者的超声心动图视频数据。每组中视频数据按约5∶1的比例分为训练集和独立测试集,采用膨胀3D卷积网络(I3D)对视频进行时空特征的联合提取,对人工智能辅助诊断模型进行训练及测试;训练集和测试集之间无病例交叉。分别使用3种不同切面[单一心尖四腔心(A4C)切面、单一胸骨旁左心室长轴(PLAX)切面、所有切面]视频数据,基于病例或视频建立模型,并进行诊断性能的统计分析,计算敏感度、特异度、受试者工作特征曲线下面积(AUC);并对比人工智能和超声科医师处理病例所需时间。结果:共纳入730例受试者,男性362例(49.6%),年龄(41.9±12.7)岁,共收集了17703条视频。其中,高血压组212例、CRF组210例、甲减组105例、正常对照组203名。使用单一PLAX切面和所有切面数据基于病例进行预测的模型诊断性能较优:(1)高血压组中,使用所有切面数据的模型的敏感度、特异度、AUC分别为97%、89%、0.93,使用单一PLAX切面模型为94%、95%、0.94;(2)CRF组中,使用所有切面数据的模型的敏感度、特异度、AUC分别为97%、95%、0.96,使用单一PLAX切面模型为97%、89%、0.93;(3)甲减组中,使用所有切面数据的模型的敏感度、特异度、AUC分别为64%、100%、0.82,使用单一PLAX切面模型为82%、89%、0.86。3D CNN模型测量和分析每例受试者的超声心动图视频所需时间明显短于超声科医师[(23.96±6.65)s比(958.25±266.17)s,P<0.001]。结论:基于3D CNN的人工智能辅助诊断模型可以联合提取超声心动图的动态时空特征,能够快速高效识别高血压性心脏病及CRF和甲减引起的心脏改变。

ObjectiveTo investigate the diagnostic efficiency and clinical application value of an artificial intelligence-assisted diagnosis model based on a three-dimensional convolutional neural network(3D CNN)on echocardiographic videos of patients with hypertensive heart disease,chronic renal failure(CRF)and hypothyroidism with cardiac involvement.MethodsThis study is a retrospective study.The patients with hypertensive heart disease,CRF and hypothyroidism with cardiac involvement,who admitted in Henan Provincial People′s Hospital from April 2019 to October 2021,were enrolled.Patients were divided into hypertension group,CRF group,and hypothyroidism group.Additionally,a simple random sampling method was used to select control healthy individuals,who underwent physical examination at the same period.The echocardiographic video data of enrolled participants were analyzed.The video data in each group was divided into a training set and an independent testing set in a ratio of 5 to 1.The temporal and spatial characteristics of videos were extracted using an inflated 3D convolutional network(I3D).The artificial intelligence assisted diagnosis model was trained and tested.There was no case overlapped between the training and validation sets.A model was established according to cases or videos based on video data from 3 different views(single apical four chamber(A4C)view,single parasternal left ventricular long-axis(PLAX)view and all views).The statistical analysis of diagnostic performance was completed to calculate sensitivity,specificity and area under the ROC curve(AUC).The time required for the artificial intelligence and ultrasound physicians to process cases was compared.ResultsA total of 730 subjects aged(41.9±12.7)years were enrolled,including 362 males(49.6%),and 17703 videos were collected.There were 212 cases in the hypertensive group,210 cases in the CRF group,105 cases in the hypothyroidism group,and 203 cases in the normal control group.The diagnostic performance of the model predicted by cases based on single PLAX view and all views data was excellent:(1)in the hypertensive group,the sensitivity,specificity and AUC of models based on all views data were 97%,89%and 0.93,respectively,while those of models based on a single PLAX view were 94%,95%,and 0.94,respectively;(2)in the CRF group,the sensitivity,specificity and AUC of models based on all views data were 97%,95%and 0.96,respectively,while those of models based on a single PLAX view were 97%,89%,and 0.93,respectively;(3)in the hypothyroidism group,the sensitivity,specificity and AUC of models based on all views data were 64%,100%and 0.82,respectively,while those of models based on a single PLAX view were 82%,89%,and 0.86,respectively.The time required for the 3D CNN model to measure and analyze the echocardiographic videos of each subject was significantly shorter than that for the ultrasound physicians((23.96±6.65)s vs.(958.25±266.17)s,P<0.001).ConclusionsThe artificial intelligence assisted diagnosis model based on 3D CNN can extract the dynamic temporal and spatial characteristics of echocardiographic videos jointly,and quickly and efficiently identify hypertensive heart disease and cardiac changes caused by CRF and hypothyroidism.