目的

针对胎儿脐膨出产前超声诊断中异常特征识别经验依赖性强的问题，开发智能辅助质控与诊断模型，以提升基层医师筛查效能。

方法

回顾性收集深圳市妇幼保健院2016 年1 月至2024 年12 月进行产前超声检查的孕妇资料，纳入脐膨出胎儿324 例1620 张图像及正常胎儿1000例2555 张图像。经超声专家标注后，基于病例水平将数据按照7：2：1 的比例划分为训练集、验证集和测试集，基于YOLO V8 构建双任务模型：脐带腹壁插入口横切面质量评估（标准/非标准）；腹壁异常结构检测（腹壁缺损、突出包块）。以精确率、召回率及平均精度评估模型检测目标结构的效能；以准确率及加权Kappa 系数分析模型切面分类与真实类别的符合率和一致性。

结果

模型对关键结构检测的平均精度达0.932，其中对腹壁轮廓（0.987）、脐带插入口（0.977）及突出包块（0.938）的检测精度表现优异，腹壁缺损检测精度稍逊（0.826）。模型对标准切面、非标准切面及脐膨出的分类准确率分别为96.8%（151/156）、100%（55/55）和96.1%（148/154），与真实类别的一致性强（加权Kappa= 0.955，P＜0.001）。

结论

基于YOLO V8 的人工智能模型实现了脐膨出超声诊断的标准化质控与异常特征识别的协同优化，尤其适用于辅助基层医师的产前超声快速筛查。

Objective

To address the issue of strong reliance on experience in identifying abnormal features during prenatal ultrasound diagnosis of fetal omphalocele， an intelligent auxiliary quality control and diagnosis model was developed to enhance the screening efficiency of primary physicians.

Methods

The data of pregnant women who underwent prenatal ultrasound examinations at Shenzhen Maternal and Child Health Hospital from January 2016 to December 2024 were retrospectively collected.A total of 324 cases of fetal omphalocele （1620 images） and 1000 cases of normal fetuses （2555 images）were included.After being labeled by ultrasound experts， the data were divided into training set， validation set， and test set at a ratio of 7：2：1 based on the case level.A dual-task model based on YOLO V8 was constructed： assessment of the quality of the cross-sectional images of the umbilical cord abdominal wall insertion （standard/non-standard） and detection of abnormal abdominal wall structures （abdominal wall defect， protruding mass）.The efficacy of the model in detecting target structures was evaluated by precision，recall rate， and average accuracy.The consistency and agreement between the model's section classification and the true category were analyzed by accuracy and weighted Kappa coefficient.

Results

The average accuracy of the model for detecting key structures reached 0.932.The detection of abdominal wall contour（0.987）， umbilical cord insertion （0.977）， and protruding mass （0.938） was excellent， while the detection accuracy of abdominal wall defect was slightly inferior （0.826）.The classification accuracy of the model for standard sections， non-standard sections， and omphalocele was 96.8% （151/156）， 100% （55/55）， and 96.1%（148/154）， respectively， and was highly consistent with the true category （weighted Kappa = 0.955， P＜0.001）.

Conclusion

The artificial intelligence model based on YOLO V8 can achieve standardized quality control and abnormal feature identification in prenatal ultrasound diagnosis of omphalocele， and is particularly suitable for assisting primary physicians in rapid prenatal ultrasound screening.