摘要目的:探讨人工智能技术在胎儿超声心动图四腔心切面中对正常、异常胎心筛查的应用价值。方法:选取北京安贞医院母胎医学研究北京市重点实验室数据库中正常和异常收缩末期四腔心切面静态图片3 996张和动态视频图像450幅作为训练集、测试集和验证集，对DGACNN模型进行训练、测试和验证。①分别选取正常、异常图片各200张，同时对本模型、DGACNN-ALOCC模型及其他达到最先进水平的分类模型(Densenet，Resnet50，InceptionV3，InceptionResnetV2)进行识别，并对结果进行比较。②将胎儿心脏超声医师分为初级、中级和高级三组，每组3人，分别选取正常、异常图片各100张，让各组医师和本模型分别对图像进行识别，每组平均分记为该组的分数，比较DGACNN模型与胎儿心脏超声学者识别的结果。结果:①假阳性率(FPR)在20%范围内时，DGACNN模型的识别准确率最高，为0.850，其他各模型识别准确率分别为DGACNN-ALOCC 0.835，Densenet 0.780，Resnet50 0.700，InceptionV3 0.670，InceptionResnetV2 0.650。②FPR在20%范围内时，DGACNN模型的ROC曲线下面积最大，为0.881，其他各模型ROC曲线下面积分别为DGACNN-ALOCC 0.864，Densenet 0.850，Resnet50 0.822，InceptionV3 0.779，InceptionResnetV2 0.703。③FPR在20%范围内时，高级胎儿心脏超声医师组的平均识别准确率最高，为0.863，其次为DGACNN模型，平均识别准确率为0.840，高于初级、中级组平均识别准确率0.760、0.807；DGACNN模型平均识别准确率高于初级、中级和高级三组的总平均识别准确率(0.810)。结论:人工智能技术在胎儿超声心动图四腔心切面筛查中是可行的，且识别准确率较高。

abstractsObjective:To investigate the value of artificial intelligence in screening normal or abnormal four-chamber view of the fetal heart.Methods:Selecting 3 996 pictures of normal and abnormal end systolic four chamber views and 450 video clips from the database of Beijing Key Laboratory of Fetal Heart Disease Maternal and Fetal Medicine Research in Beijing Anzhen Hospital as training set, test set and verification set to train, test and verify DGACNN model. ①Comparing DGACNN, DGACNN-ALOCC and other classification models(Densenet, Resnet50, InceptionV3, InceptionResnetV2) to detect the model with the most advanced level by recognizing 200 normal pictures and 200 abnormal pictures. ②Fetal echocardiographers were divided into three groups according to their experiences: primary, intermediate and advanced, 3 doctors in each group, and comparing the average score between each group or three groups and DGACNN by recognizing 100 normal pictures and 100 abnormal pictures.Results:①When the the false positive rate(FPR) was in the range of 20%, the recognition accuracy of DGACNN was the highest with 0.850, the recognition accuracy of other models were DGACNN-ALOCC 0.835, Densenet 0.780, Resnet50 0.700, InceptionV3 0.670, InceptionResnetV2 0.650, respectively. ②When FPR was in the range of 20%, the area under ROC curve of DGACNN was the largest with 0.881, the area under ROC curve of other models were DGACNN-ALOCC 0.864, Densenet 0.850, Resnet50 0.822, Inceptionv3 0.779, InceptionResnetV2 0.703, respectively. ③When the FPR was in the range of 20%, the average recognition accuracy of the senior fetal echocardiographer group was the highest with 0.863, followed by DGACNN 0.840, which was higher than the average recognition accuracy of the primary and intermediate groups with 0.760, 0.807; the average recognition accuracy of DGACNN was higher than the total average recognition accuracy of the primary, intermediate and advanced groups with 0.810.Conclusions:Artificial intelligence is accessible in screening four chamber view of fetal echocardiography, with high recognition accuracy.