早孕期胎儿头臀长正中矢状切面超声图像的人工智能质控研究

Artificial intelligence-based quality control of mid-sagittal plane ultrasound images for first trimester fetal crown-rump length

目的探讨人工智能对早孕期胎儿头臀长正中矢状切面超声图像标准程度判断的临床应用价值。方法选取深圳市妇幼保健院2022年1月至12月11~13^(+6)周早孕期胎儿头臀长正中矢状切面超声图像1251张为研究对象,以产前超声专家委员会对图像标准程度判断的结果作为金标准,对比智能质控模型、高级超声医师、中级超声医师、初级超声医师对图像标准程度判断的符合率,应用McNemar-Bowker检验和Weighted Kappa分析组间结果的差异性与一致性;以每100张图像为1组,记录每组图像的质控耗时,应用两相关样本Wilcoxon符号秩检验比较四者质控耗时的差异。结果智能质控模型对胎儿头臀长正中矢状切面超声图像标准程度判断符合率为90.5%,与金标准结果一致性强(Kappa值=0.83,P<0.001),略低于高级超声医师(91.1%),差异具有统计学意义(χ^(2)=16.40,P<0.001),优于中级超声医师(78.7%)和初级超声医师(68.9%),差异具有统计学意义(χ^(2)=100.25、16.88,P均<0.001)。智能质控模型每组超声图像质控耗时明显少于超声医师[3.57(3.55,3.60)s vs 351(309,384)s vs 363(351,370)s vs 433(407,475)s],差异具有统计学意义(Z=-3.180、-3.181、-3.180、P均<0.001)。结论智能质控模型对早孕期胎儿头臀长正中矢状切面超声图像标准程度判断准确且高效。

Objective To probe into the clinical application value of artificial intelligence in the judgment of the quality standard of mid-sagittal plane ultrasound images for first trimester fetal crown-rump length(CRL).Methods A total of 1251 midsagittal plane ultrasound images of fetuese at 11-13^(+6)weeks of gestation were selected from the database of Shenzhen Maternity&Child Healthcare Hospital from January to December 2022.Using the unified judging results of the image quality standard by the Expert Committee of Prenatal Ultrasound as the golden standard,the performance of an artificial intelligent based quality control model,senior,middle,and junior sonographers in the judgment of the quality standard of mid-sagittal plane ultrasound images for first trimester CRL was assessed by calculating their coincidence rates with the golden standard.The coincidence rates were compared using the Mcnemar-Bowker tests,and weighted Kappa values were applied to analyse the difference and consistency among these results.Time of quality control for each set of images was recorded as one group per 100 images.The Wilcoxon's two samples signed-rank test was applied to compare the difference in the time spent among the four groups.Results The coincidence rate of the intelligent quality control model with the golden standard was 90.5%,suggesting a strong consistency to the golden standard(Kappa=0.83,P<0.001),which was slightly lower than that of senior sonographers(90.5%vs 91.1%,χ^(2)=16.40,P<0.001),but superior to that of middle and junior sonographers'(90.5%vs 78.7%vs 68.9%,χ^(2)=100.25,16.88,P<0.001 for all).The time spent by the intelligent model quality control was significantly less than that by ultrasound physicians[3.57(3.55,3.60)s vs 351(309,384)s vs 363(351,370)s vs 433(407,475)s;Z=-3.180,-3.181,and-3.180,respectively,P<0.001 for all].Conclusion The intelligent quality control model is accurate and efficient in the judgment of the quality standard of mid-sagittal plane ultrasound images for first trimester CRL.