2025 , Vol. 22 >Issue 05: 388 - 396
DOI: https://doi.org/10.3877/cma.j.issn.1672-6448.2025.05.002
人工智能在胎儿超声心动图标准切面质量控制中的多中心应用研究
Copy editor: 吴春凤
收稿日期: 2025-04-11
网络出版日期: 2025-07-17
版权
Application of artificial intelligence in quality control of standard views for fetal echocardiography: a multi-center study
Received date: 2025-04-11
Online published: 2025-07-17
Copyright
目的
探讨人工智能深度学习算法在胎儿心脏11 个标准切面质量控制中的应用价值。
方法
选择7 家医院胎儿超声心动图检查病例图片,筛选其中妊娠20 ~34 周的胎儿心脏图像共35 331 张。根据胎儿心脏检查超声指南推荐需要留存的11 个胎儿心脏切面,提出一种融合Transformer 技术(深度学习的架构之一)的心脏切面自动质量控制方法对胎儿心脏图像进行质量控制。以专家团队评定的结果为标准,将收集到的胎儿心脏图像分为数据集A 和数据集B,分别包含24 000、11 331 张图像。数据集A 用于深度学习模型训练,数据集B 则分别用于深度学习预测以及2名具有5 年超声心动图检查经验的医师进行人工质量评估。采用平均精度(AP)值作为核心统计指标评估模型性能。
结果
融合Transformer 技术的心脏切面自动质量控制方法对于胎儿心脏切面解剖结构识别的AP 值达到了0.885,可以准确识别胎儿超声心动图需要留存切面常用解剖结构。深度学习模型识别胎儿心脏超声图像平均时长为0.028 s/张,而2 名超声医师识别的平均时长为3.77 s/张,使用基于深度学习的方法评估胎儿心脏超声图像质量的速度是医师人工评估的134.6 倍。
结论
运用深度学习模型对胎儿心脏超声切面进行质量控制可以达到专家判定的水平并节省了时间。
何冠南 , 谭莹 , 路玉欢 , 蒲斌 , 扬水华 , 张仁铁 , 陈明 , 石智红 , 钟晓红 , 陈曦 , 燕柳屹 , 李胜利 . 人工智能在胎儿超声心动图标准切面质量控制中的多中心应用研究[J]. 中华医学超声杂志(电子版), 2025 , 22(05) : 388 -396 . DOI: 10.3877/cma.j.issn.1672-6448.2025.05.002
Objective
To explore the application value of deep learning algorithms in quality control of the 11 standard fetal cardiac views.
Methods
Images of fetal echocardiography were collected from seven hospitals, of which a total of 35331 images were selected from fetuses between 20 and 34 weeks of gestation.Based on the 11 standard fetal cardiac views recommended by fetal echocardiography guidelines, a novel automatic quality control method integrating transformer-based techniques was proposed to assess image quality.Using expert evaluation as the reference standard, the collected images were divided into two datasets: dataset A (24 000 images) for model training, and dataset B (11 331 images) for both deep learning-based prediction and manual quality assessment by two physicians with five years of clinical experience.Average precision (AP) was used as the primary metric to evaluate model performance.
Results
The transformer-based automatic quality control method achieved an AP of 0.885 in recognizing anatomical structures in fetal cardiac views, demonstrating accurate identification of key anatomical features required in standard fetal echocardiography.The deep learning model processed each image in approximately 0.028 seconds, while the two experienced physicians took an average of 3.77 seconds per image.Thus, the deep learning-based approach was 134.6 times faster than manual evaluation.
Conclusion
The application of deep learning models for quality control of fetal echocardiographic views can achieve expert-level performance while significantly reducing the time required for manual quality assessment.
Key words: Artificial intelligence; Fetal cardiac; Quality control
表1 妊娠20 ~34 周正常胎儿超声心动图数据集分布 |
切面类型 | 数据集 A | 数据集 B | |||||
---|---|---|---|---|---|---|---|
训练集 | 验证集 | 测试集 | 合计 | 标准切面 | 非标准切面 | 合计 | |
DuArch | 1200 | 400 | 400 | 2000 | 426 | 425 | 851 |
3VT | 1202 | 399 | 399 | 2000 | 480 | 467 | 947 |
3VV | 1200 | 400 | 400 | 2000 | 427 | 374 | 801 |
AC | 1200 | 400 | 400 | 2000 | 402 | 375 | 777 |
CLA | 1202 | 399 | 399 | 2000 | 447 | 388 | 835 |
4CH | 1200 | 400 | 400 | 2000 | 581 | 618 | 1199 |
SA | 1202 | 399 | 399 | 2000 | 378 | 284 | 662 |
RVOT | 1204 | 398 | 398 | 2000 | 536 | 293 | 829 |
LVOT | 1202 | 399 | 399 | 2000 | 458 | 329 | 787 |
LRPA | 1200 | 400 | 400 | 2000 | 374 | 358 | 732 |
AoArch | 1200 | 400 | 400 | 2000 | 423 | 412 | 835 |
其他切面 | 1200 | 400 | 400 | 2000 | - | - | 2076 |
合计 | 14412 | 4794 | 4794 | 24000 | - | - | 11331 |
注:DuArch 为动脉导管弓切面,3VT 为三血管气管切面,3VV 为三血管切面,AC 为腹部横切面,CLA 为上、下腔静脉长轴切面,4CH 为四腔心切面,SA 为大动脉短轴切面,RVOT 为右心室流出道切面,LVOT 为左心室流出道切面,LRPA 为左、右肺动脉分叉切面,AoArch 为主动脉弓切面;-表示无对应数据或无法合计 |
图1 人工智能模型检测胎儿超声心动图11 个标准切面上解剖结构框及标准切面示意图注:左边为超声切面图像,右边为该切面的模式图(Schematic),从左到右依次为AC(腹部横切面)、4CH(四腔心切面)、LVOT(左心室流出道切面)、RVOT(右心室流出道切面、SA(大动脉短轴切面)、3VV(三血管切面)、LRPA(左、右肺动脉分叉切面)、3VT(三血管气管切面)、AoArch(主动脉弓切面)、DuArch(动脉导管弓切面)、CLA(上、下腔静脉长轴切面);R 为右侧,L 为左侧,RIB 为肋骨,IVC 为下腔静脉,AW 为腹壁,UV 为脐静脉,PV 为门静脉,SP 为脊柱,ST 为胃,AO 为主动脉,LV 为左心室,RV 为右心室,LA 为左心房,RA 为右心房,DAO 为降主动脉,IVS 为室间隔,LVTO 为左心室流出道,ARCH 为主动脉弓,SVC 为上腔静脉,MPA 为主肺动脉,RPA 为右肺动脉,AAO 为升主动脉,ICSA 为头颈部血管分支 |
表2 本研究模型算法与同类算法在胎儿超声心动图切面类型和解剖结构识别准确率方面比较的结果 |
项 目 | YOLOF | RSB | DDOD | Anchor Pruning | DyHead | SABL | AutoAssign | CentripetalNet | ATSS | FoveaBox | 心脏智能质控模型 |
---|---|---|---|---|---|---|---|---|---|---|---|
切面 | |||||||||||
AP50 | 0.839 | 0.839 | 0.744 | 0.557 | 0.822 | 0.854 | 0.848 | 0.814 | 0.835 | 0.829 | 0.877 |
DuArch | 0.889 | 0.902 | 0.822 | 0.677 | 0.879 | 0.895 | 0.913 | 0.866 | 0.886 | 0.892 | 0.911 |
3VT | 0.724 | 0.760 | 0.661 | 0.348 | 0.679 | 0.768 | 0.760 | 0.639 | 0.717 | 0.729 | 0.789 |
3VV | 0.665 | 0.658 | 0.557 | 0.264 | 0.606 | 0.694 | 0.677 | 0.554 | 0.654 | 0.659 | 0.704 |
AC | 0.999 | 0.999 | 0.897 | 0.981 | 0.991 | 0.999 | 1.000 | 0.954 | 0.981 | 0.946 | 1.000 |
CLA | 0.815 | 0.870 | 0.793 | 0.602 | 0.832 | 0.876 | 0.845 | 0.826 | 0.838 | 0.830 | 0.849 |
4CH | 0.966 | 0.965 | 0.903 | 0.889 | 0.962 | 0.965 | 0.965 | 0.893 | 0.960 | 0.963 | 0.972 |
SA | 0.799 | 0.793 | 0.750 | 0.503 | 0.786 | 0.791 | 0.813 | 0.717 | 0.789 | 0.772 | 0.831 |
RVOT | 0.649 | 0.639 | 0.551 | 0.225 | 0.590 | 0.662 | 0.655 | 0.532 | 0.653 | 0.649 | 0.699 |
LVOT | 0.803 | 0.778 | 0.740 | 0.372 | 0.686 | 0.813 | 0.791 | 0.705 | 0.739 | 0.765 | 0.830 |
LRPA | 0.854 | 0.853 | 0.814 | 0.266 | 0.837 | 0.874 | 0.850 | 0.811 | 0.865 | 0.848 | 0.875 |
AoArch | 0.855 | 0.872 | 0.790 | 0.693 | 0.852 | 0.899 | 0.886 | 0.861 | 0.858 | 0.871 | 0.904 |
其他切面 | 0.959 | 0.949 | 0.538 | 0.906 | 0.961 | 0.956 | 0.966 | 0.669 | 0.960 | 0.888 | 0.972 |
切面框AP | 0.831 | 0.837 | 0.735 | 0.561 | 0.805 | 0.849 | 0.843 | 0.752 | 0.825 | 0.818 | 0.861 |
解剖结构 | |||||||||||
AAO | 0.894 | 0.899 | 0.811 | 0.704 | 0.881 | 0.893 | 0.905 | 0.913 | 0.885 | 0.876 | 0.922 |
AO | 0.884 | 0.916 | 0.682 | 0.531 | 0.907 | 0.895 | 0.917 | 0.942 | 0.917 | 0.870 | 0.930 |
ARCH | 0.859 | 0.863 | 0.791 | 0.520 | 0.852 | 0.873 | 0.864 | 0.881 | 0.868 | 0.859 | 0.895 |
AW | 0.997 | 0.999 | 0.985 | 0.983 | 0.989 | 0.962 | 0.999 | 0.974 | 0.989 | 0.998 | 1.000 |
DAO | 0.925 | 0.915 | 0.853 | 0.743 | 0.916 | 0.933 | 0.924 | 0.919 | 0.923 | 0.914 | 0.946 |
IAVS | 0.888 | 0.868 | 0.714 | 0.540 | 0.855 | 0.880 | 0.877 | 0.859 | 0.875 | 0.890 | 0.922 |
ICSA | 0.587 | 0.592 | 0.521 | 0.149 | 0.549 | 0.662 | 0.660 | 0.520 | 0.567 | 0.601 | 0.746 |
IVC | 0.811 | 0.786 | 0.654 | 0.349 | 0.773 | 0.798 | 0.810 | 0.815 | 0.814 | 0.748 | 0.817 |
IVS | 0.827 | 0.855 | 0.741 | 0.412 | 0.798 | 0.836 | 0.839 | 0.855 | 0.809 | 0.808 | 0.892 |
LA | 0.844 | 0.818 | 0.755 | 0.484 | 0.801 | 0.846 | 0.831 | 0.850 | 0.825 | 0.851 | 0.876 |
LAAO | 0.765 | 0.745 | 0.636 | 0.229 | 0.723 | 0.774 | 0.754 | 0.790 | 0.732 | 0.728 | 0.806 |
LPA | 0.787 | 0.801 | 0.722 | 0.538 | 0.817 | 0.841 | 0.806 | 0.808 | 0.815 | 0.798 | 0.849 |
LV | 0.881 | 0.866 | 0.756 | 0.505 | 0.834 | 0.888 | 0.878 | 0.870 | 0.852 | 0.852 | 0.917 |
MPA | 0.804 | 0.792 | 0.728 | 0.387 | 0.803 | 0.852 | 0.810 | 0.833 | 0.805 | 0.843 | 0.861 |
MPADA | 0.847 | 0.841 | 0.765 | 0.618 | 0.834 | 0.856 | 0.840 | 0.821 | 0.832 | 0.838 | 0.854 |
RA | 0.882 | 0.870 | 0.805 | 0.639 | 0.864 | 0.889 | 0.877 | 0.858 | 0.871 | 0.869 | 0.904 |
RIB | 0.787 | 0.702 | 0.657 | 0.530 | 0.758 | 0.782 | 0.753 | 0.664 | 0.771 | 0.739 | 0.812 |
RPA | 0.854 | 0.885 | 0.769 | 0.583 | 0.844 | 0.860 | 0.855 | 0.871 | 0.858 | 0.840 | 0.906 |
RV | 0.838 | 0.822 | 0.720 | 0.513 | 0.818 | 0.829 | 0.824 | 0.810 | 0.809 | 0.809 | 0.866 |
SP | 0.955 | 0.945 | 0.911 | 0.931 | 0.951 | 0.946 | 0.952 | 0.937 | 0.951 | 0.939 | 0.956 |
ST | 0.974 | 0.969 | 0.951 | 0.939 | 0.971 | 0.980 | 0.974 | 0.969 | 0.977 | 0.971 | 0.973 |
SVC | 0.768 | 0.778 | 0.642 | 0.383 | 0.753 | 0.803 | 0.795 | 0.821 | 0.783 | 0.770 | 0.844 |
T | 0.611 | 0.698 | 0.464 | 0.221 | 0.662 | 0.684 | 0.708 | 0.741 | 0.682 | 0.654 | 0.779 |
UVPV | 0.966 | 0.953 | 0.917 | 0.892 | 0.962 | 0.972 | 0.964 | 0.955 | 0.966 | 0.950 | 0.972 |
结构框AP | 0.843 | 0.841 | 0.748 | 0.555 | 0.830 | 0.856 | 0.851 | 0.845 | 0.841 | 0.834 | 0.885 |
注:AP 为平均精度值,AP50 表示切面类别框与解剖结构框的平均准确率,切面框AP 表示切面框的平均准确率,结构框AP 表示解剖结构框的平均准确率。DuArch 为动脉导管弓切面,3VT 为三血管气管切面,3VV 为三血管切面,AC 为腹部横切面,CLA 为上、下腔静脉长轴切面,4CH 为四腔心切面,SA 为大动脉短轴切面,RVOT 为右心室流出道切面,LVOT 为左心室流出道切面,LRPA 为左、右肺动脉分叉切面,AoArch为主动脉弓切面;AAO 为升主动脉,AO 为主动脉,ARCH 为主动脉弓,AW 为腹壁,DAO 为降主动脉,IAVS 为房室间隔,ICSA 为头颈部血管分支,IVC 为下腔静脉,IVS 为室间隔,LA 为左心房,LAAO 为左房及主动脉,LPA 为左肺动脉,LV 为左心室,MPA 为主肺动脉,MPADA 为主肺动脉及动脉导管,RA 为右心房,RIB 为肋骨,RPA 为右肺动脉,RV 为右心室,SP 为脊柱,ST 为胃,SVC 为上腔静脉,T 为气管,UVPU 为脐静脉及门静脉;YOLOF、RSB、DDOD、Anchor Pruning、DyHead、SABL、AutoAssign、CentripetalNet、ATSS、FoveaBox 为10 种先进检测算法 |
表3 心脏智能质控模型及医师组对胎儿超声心动图切面类型及切面标准性判断结果比较 |
切面类别 | 标准与否 | 切面标准与否准确性 | 切面类型准确性 | ||||
---|---|---|---|---|---|---|---|
医师1 | 医师2 | 心脏智能质控模型 | 医师1 | 医师2 | 心脏智能质控模型 | ||
DuArch | 标准 | 0.881 | 0.785 | 0.718 | 0.542 | 0.179 | 0.780 |
非标准 | 0.427 | 0.600 | 0.870 | ||||
3VT | 标准 | 0.998 | 0.788 | 0.876 | 0.624 | 0.301 | 0.630 |
非标准 | 0.026 | 0.382 | 0.660 | ||||
3VV | 标准 | 0.988 | 0.392 | 0.832 | 0.546 | 0.397 | 0.551 |
非标准 | 0.070 | 0.857 | 0.611 | ||||
AC | 标准 | 0.832 | 0.857 | 0.930 | 0.961 | 0.737 | 0.981 |
非标准 | 0.859 | 0.651 | 0.793 | ||||
CLA | 标准 | 0.932 | 0.788 | 0.843 | 0.564 | 0.551 | 0.674 |
非标准 | 0.343 | 0.636 | 0.828 | ||||
4CH | 标准 | 0.921 | 0.838 | 0.835 | 0.827 | 0.490 | 0.823 |
非标准 | 0.341 | 0.570 | 0.627 | ||||
SA | 标准 | 0.952 | 0.532 | 0.717 | 0.834 | 0.471 | 0.801 |
非标准 | 0.114 | 0.730 | 0.816 | ||||
RVOT | 标准 | 0.983 | 0.330 | 0.877 | 0.264 | 0.211 | 0.604 |
非标准 | 0.071 | 0.722 | 0.491 | ||||
LVOT | 标准 | 0.962 | 0.735 | 0.896 | 0.644 | 0.363 | 0.738 |
非标准 | 0.071 | 0.586 | 0.304 | ||||
LRPA | 标准 | 0.968 | 0.723 | 0.912 | 0.478 | 0.403 | 0.652 |
非标准 | 0.287 | 0.754 | 0.678 | ||||
AoArch | 标准 | 0.920 | 0.936 | 0.691 | 0.733 | 0.430 | 0.784 |
非标准 | 0.362 | 0.360 | 0.795 | ||||
其他切面 | 0.895 | 0.948 | 0.711 | ||||
平均值 | 0.605 | 0.662 | 0.755 | 0.659 | 0.457 | 0.727 |
注:DuArch 为动脉导管弓切面,3VT 为三血管气管切面,3VV 为三血管切面,AC 为腹部横切面,CLA 为上、下腔静脉长轴切面,4CH 为四腔心切面,SA 为大动脉短轴切面,RVOT 为右心室流出道切面,AoArch 为主动脉弓切面,LVOT 为左心室流出道切面,LRPA 为左、右肺动脉分叉切面 |
表4 心脏智能质控模型及医师组判断胎儿超声心动图切面类型检测速度结果(s/张) |
切面类型 | 医师1 | 医师2 | 心脏智能质控模型 |
---|---|---|---|
DuArch | 5.944 | 2.401 | 0.028 |
3VT | 4.600 | 2.810 | |
3VV | 5.811 | 3.090 | |
AC | 4.314 | 2.918 | |
CLA | 4.775 | 3.073 | |
4CH | 4.178 | 2.430 | |
SA | 5.264 | 3.740 | |
RVOT | 5.302 | 3.170 | |
LVOT | 4.140 | 2.638 | |
LRPA | 5.559 | 3.999 | |
AoArch | 5.873 | 2.674 | |
其他切面 | 2.873 | 1.892 | |
平均值 | 4.886 | 2.903 |
注:DuArch 为动脉导管弓切面,3VT 为三血管气管切面,3VV 为三血管切面,AC 为腹部横切面,CLA 为上、下腔静脉长轴切面,4CH 为四腔心切面,SA 为大动脉短轴切面,RVOT 为右心室流出道切面,AoArch 为主动脉弓切面,LVOT 为左心室流出道切面,LRPA 为左、右肺动脉分叉切面 |
1 |
Garcia-Canadilla P, Sanchez-Martinez S, Crispi F, et al.Machine learning in fetal cardiology: what to expect [J].Fetal Diagn Ther,2020, 47(5): 363-372.
|
2 |
Gómez Montes E, Herraiz I, Mendoza A, et al.Fetal intervention in right outflow tract obstructive disease: selection of candidates and results [J].Cardiol Res Pract, 2012, 2012: 592403.
|
3 |
宋书豪, 曾施.应用深度学习模型分类正常胎儿心脏超声切面 [J].中国医学影像技术, 2025, 41(1): 70-73.
|
4 |
陈雯雯, 朱业, 张易薇, 等.深度学习模型用于自动分类超声心动图切面 [J].中国医学影像技术, 2024, 40(8): 1124-1129.
|
5 |
成汉林, 史中青, 戚占如, 等.基于深度学习的超声心动图动态图像切面识别研究 [J/OL].中华医学超声杂志(电子版), 2024, 21(2):128-136.
|
6 |
石智红, 李胜利.人工智能在产前超声中的应用和研究进展 [J/OL].中华医学超声杂志(电子版), 2023, 20(1): 113-117.
|
7 |
李胜利, 秦越, 谭光华, 等.医学超声人工智能的应用与挑战 [J/OL].中华医学超声杂志(电子版), 2023, 20(1): 1-5.
|
8 |
Pu B, Wang L, Yang J, et al.M3-UDA: a new benchmark for unsupervised domain adaptive fetal cardiac structure detection [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2024: 11621-11630.
|
9 |
Wu L, Cheng JZ, Li S, et al.FUIQA: fetal ultrasound image quality assessment with deep convolutional networks [J].IEEE Trans Cybern,2017, 47(5): 1336-1349.
|
10 |
Abdi AH, Luong C, Tsang T, et al.Automatic quality assessment of echocardiograms using convolutional neural networks: feasibility on the apical four-chamber view [J].IEEE Trans Med Imaging, 2017,36(6): 1221-1230.
|
11 |
Abdi AH, Luong C, Tsang T, et al.Quality assessment of echocardiographic cine using recurrent neural networks: feasibility on five standard view planes [C]//Medical Image Computing and Computer Assisted Intervention - MICCAI 2017.Cham, 2017: 302-310.
|
12 |
Pu B, Li K, Li S, et al.Automatic fetal ultrasound standard plane recognition based on deep learning and IIoT [J].IEEE Trans Industr Inform, 2021, 17(11): 7771-7780.
|
13 |
Pu B, Lu Y, Chen J, et al.Mobileunet-fpn: a semantic segmentation model for fetal ultrasound four-chamber segmentation in edge computing environments [J].IEEE J Biomed Health Inform, 2022,26(11): 5540-5550.
|
14 |
Pu B, Zhu N, Li K, et al.Fetal cardiac cycle detection in multiresource echocardiograms using hybrid classification framework [J].Future Generation Computer Systems, 2021, 115: 825-836.
|
15 |
Lu Y, Li K, Pu B, et al.A YOLOX-based Deep Instance Segmentation Neural Network for Cardiac Anatomical Structures in Fetal Ultrasound Images [J].IEEE/ACM Trans Comput Biol Bioinform, 2022, 21(4):1007-1018.
|
16 |
Shan A, Lv J, Zhu H, et al.Fetal heart and descending aorta detection in four-chamber view of fetal echocardiography [J].Annu Int Conf IEEE Eng Med Biol Soc, 2021, 2021: 2722-2725.
|
17 |
American Institute of Ultrasound in Medicine.AIUM practice guideline for the performance of fetal echocardiography [J].J Ultrasound Med, 2013, 32(6): 1067-1082.
|
18 |
International Society of Ultrasound in Obstetrics and Gynecology,Carvalho JS, Allan LD, et al.ISUOG Practice Guidelines (updated):sonographic screening examination of the fetal heart [J].Ultrasound Obstet Gynecol, 2013, 41(3): 348-359.
|
19 |
Chinese standard for fetal echocardiography [J].Chin J Ultrasonogr,2015, 24(11): 921-926.
|
20 |
Abdi AH, Luong C, Tsang T, et al.Automatic quality assessment of echocardiograms using convolutional neural networks: feasibility on the apical four-chamber view [J].IEEE Trans Med Imaging, 2017,36(6): 1221-1230.
|
21 |
Dong J, Liu S, Liao Y, et al.A generic quality control framework for fetal ultrasound cardiac four-chamber planes [J].IEEE J Biomed Health Inform, 2020, 24(4): 931-942.
|
/
〈 |
|
〉 |