Objective:To develop an artificial intelligence (AI) quality control model of fetal heart in the first trimester and verify its effectiveness.Methods:A total of 18 694 images of the four-chamber view(4CV) and three-vessel and tracheal view(3VT) of fetal heart in the first trimester were selected from Shenzhen Maternal and Child Health Hospital Affiliated to Southern Medical University since January 2022 to December 2022. A total of 14 432 images were manually annotated. The one-stage target detection algorithm YOLO V5 was used to train the AI quality control model in the first trimester of fetal heart, and 4 262 images (golden standard set by expert group) were used to evaluate the application effectiveness of AI quality control model. Kappa consistency test was used to compare the results of section classification and standard degree judgment from AI quality control model, Doctor 1(D1) and Doctor 2(D2).Results:①Precision of the AI quality control model was 0.895, recall was 0.852, mean average precision (mAP 50) was 0.873.The average precision(AP) of the AI quality control model for section classification was 0.907 (4CV) and 0.989 (3VT), respectively. ②Compared with the gold standard, the overall coincidence rate and consistency of section classification of AI quality control model, D1 and D2 were 99.91% (Kappa=0.998), 100% (Kappa=1.000), 100% (Kappa=1.000), respectively. The coincidence rate and consistency of the plane standard degree evaluation from the AI quality control model, D1 and D2 were 97.46% (Weighted Kappa=0.932), 93.73% (Weighted Kappa=0.847), and 93.12% (Weighted Kappa=0.832), respectively. Strong consistency was displayed. Moreover, AI quality control model showed the highest coincidence rate and the strongest consistency in judging section standard degree, which was superior to manual quality control. The time-consuming of AI quality control (0.012 s/sheet) was significantly less than the way of manual quality control (4.76-6.11 s/sheet)( Z=-8.079, P<0.001). Conclusions:The use of artificial intelligent fetal heart quality control model in the first trimester can effectively and accurately control the image quality.

Objective:To explore the effectiveness of artificial intelligent (AI) quality control system on the standard mid-sagittal view of fetal face in 11-13 +6 weeks of gestation. Methods:The quality of the images was evaluated using online " Intelligent ultrasonic quality control system" for 1 063 sections of fetal Nuchal translucency ultrasound images in the Ultrasound Department of Shenzhen Maternity and Child Healthcare Hospital of Southern Medical University. The manual quality control results and time consuming of two experts were compared with AI. The gold standard was decided by experts group with higher seniority.Results:The overall standard rate, substandard rate and nonstandard rate of the image evaluated by the intelligent quality control system was 87.3%, 2.82%, and 9.88%. The overall accuracy of images was 96.6%. The coincidence rates between AI and the two experts were 96.1% and 96.3%, respectively, with strong consistency (Kappa were 0.835 and 0.845 respectively). The time required for intelligent quality control was significantly shorter than manual quality control (208 s vs 6 696 s/6 602 s). All the differences were statistically significant ( Z=-3.981, P<0.001). Conclusions:The intelligent quality control system could accurately and quickly evaluate whether the mid-sagittal view of fetal face in 11-13 +6 weeks of gestation is standard.

目的 观察产科超声图像智能质量控制系统(IU QCS)的效能.方法 以IU QCS评价64家深圳医院573名医师采集的164010幅6774胎孕中晚期胎儿声像图的质量,由2名专家(E1/E2)对其中57444幅单幅图像进行质控;对比其对切面分类及标准程度的质控效能和一致性及耗时差异.结果 IU QCS质控总体标准率为81.16％,基本标准率10.10％,非标准率8.74％.IU QCS与E1/E2质控切面分类总体符合率分别为97.61％(56071/57444)及97.65％(56092/57444),一致性强(Kappa均>0.933);标准程度评价总体符合率为89.72％(50307/56071)及89.67％(50300/56092),一致性较强(Kappa均=0.658).IUQCS每100幅质控耗时明显小于人工质控[33(29,37)s vs.705(680,730)s vs.720(696,751)s,Z均=-20.776,P均<0.001].结论 利用产科IUQCS可较为准确、高效地实现图像质控.

Objective:To assess the clinical value of online artificial intelligent quality control system in assessing obstetric ultrasound images.Methods:A total of 374 191 images of 15 640 obstetric ultrasound cases from 998 doctors in 137 hospitals in Shenzhen and Chongqing were selected from January 1 to June 30, 2021, and the quality of the images was evaluated using online artificial intelligent quality control system. Based on each image quality assessment result, the proportion of different standard levels of all planes was calculated. The appeal results were also recorded to observe the accuracy of the system. To survey the efficiency of the system, paired sample t test was used to compare the time spent by intelligent quality control and manual quality control.Results:The overall standard rate, substandard rate, and nonstandard rate were 81.18%, 12.06%, and 6.76%, respectively. A total of 285 appealed images (0.076%) were reviewed by authoritative experts, who confirmed the initial diagnosis in 126 images (44.21%), and did not support the initial diagnosis in 159 images (55.79%); the accuracy of the system reached 99.96% (374 032/374 191). The average time spent by intelligent quality control for 100 images was (32.7±5.1) s, significantly shorter than that spent by manual quality control by two ultrasound physicians [(705.3±37.2) s and (724.6±40.4) s, t=62.667 and 56.396, respectively, P＜0.001].Conclusion:The intelligent quality control system of obstetric ultrasound images allows the quality control to be performed objectively, accurately, and efficiently, which is of great significance to guide the improvement of image quality.

目的 探讨人工智能(artificial intelligence,AI)质量控制系统在提高胎儿上腹部水平横切面超声图像标准率中的应用价值.方法 应用"产前超声AI智慧云平台"对深圳市60家医院三个季度在产科超声检查中存储的胎儿上腹部水平横切面图像共18114张进行智能质控,受质控医生通过查阅质控结果及图像存在的不足之处针对性地加以改进.应用χ2检验对三个季度图像的标准率、不足原因及申诉情况进行两两比较,评估智能质控对于提高胎儿上腹部水平横切面标准程度的价值.结果 各医院第一、二、三季度胎儿上腹部水平横切面图像标准率分别为80.15％(5649/7048)、86.2％(4391/5096)、90.55％(5406/5970),各季度图像标准率逐步提高,两两比较差异有统计学意义(P<0.01);每个季度的不标准图像中,图像不足的主要原因是切面中出现了大片肺脏,分别为8.9％(631/7048)、7.0％(358/5096)、4.9％(294/5970),其次是脐静脉和门静脉汇合部显示不清,分别为4.3％(305/7048)、3.5％(181/5096)、2.6％(155/5970),这两类图像均显著减少,两两季度比较差异有统计学意义(P<0.01);医生申诉后维持智能评价结果的图像分别为66.4％(79/119)、49％(25/51)、33.3％(10/30),两两比较差异有统计学意义(P<0.01),提示各季度医生对图像认识错误率逐渐降低.结论 人工智能质量控制可有效提高胎儿上腹部水平横切面超声图像的标准率,提升医生对标准切面的认识.