基于实时目标检测网络的胎儿颜面部超声切面识别及应用

Fetal facial ultrasound plane recognition based on real-time object detection network and its application

目的:探讨基于实时目标检测网络的人工智能(AI)模型在胎儿颜面部超声检查中的应用价值。方法:以妊娠20~24周正常胎儿颜面部超声标准切面(FFUSP)图像为研究对象,构建基于实时目标检测网络的FFUSP识别模型,观察其对FFUSP及其解剖结构的识别精度;通过临床验证分析其对119例胎儿超声图像中FFUSP识别效能以评价其临床应用价值。结果:AI模型对胎儿颜面部结构识别的整体查准率为97.8%、查全率为98.5%、mAP@.5为98.1%、mAP@.5:.95为61.0%。在临床验证中,AI模型对颜面部解剖结构识别的敏感度、特异度、阳性预测值、阴性预测值及准确率分别为100.0%、98.5%、87.4%、100.0%、98.7%,与胎儿超声专家分类一致性强(k=0.925,P<0.001);对3类标准切面图像的识别准确率为100%;动态视频检测平均速度为33.93帧/s。结论:基于实时目标检测网络的FFUSP识别模型性能优越,可应用于实时超声检查辅助诊断、教学及智能化质量评价。

Objective To explore the role of an artificial intelligence(AI) model based on real-time object detection network in fetal facial ultrasound examination.Methods With the normal fetal facial ultrasound standard plane(FFUSP) at 20-24 weeks of gestation as the research object,a FFUSP recognition model based on real-time object detection network was constructed.The recognition accuracy of the model for FFUSP and the anatomical structures were analyzed,and the clinical value was evaluated by analyzing its performance in identifying FFUSP in 119 cases of fetal ultrasound images.Results The overall precision,recall rate,mAP@.5 and mAP@.5:.95 of the AI model were 97.8%,98.5%,98.1% and 61.0%,respectively.The clinical validation showed that the AI model had a sensitivity,specificity,positive predictive value,negative predictive value and accuracy of 100.0%,98.5%,87.4%,100.0% and 98.7% for facial anatomy recognition,and the results were highly consistent with the classification of fetal ultrasound experts(k=0.925,P<0.001).The recognition accuracy of the model for 3types of standard planes reached 100%;and the average speed of dynamic video detection was 33.93 frames per second.Conclusion The FFUSP recognition model based on real-time object detection network exhibits excellent performance,and it can be applied to real-time ultrasound diagnosis,teaching and intelligent quality evaluation.