探讨人工智能技术在小儿先天性心脏病超声筛查中的应用价值。

选取2021年9月至2022年2月于浙江大学医学院附属儿童医院进行超声心动图检查的心脏结构正常及异常的彩色多普勒超声静态图像8543张，按8∶1∶1的比例将其划分为训练集6871张、验证集833张、测试集839张。构建OBICnet模型，采用F1分数、准确率、精确率、召回率、特异度、误诊率及漏诊率评价人工智能模型的识别性能，并与ResNet50、GBCnet两种人工智能模型进行比较。另收集2022年11月至2023年1月在浙江大学医学院附属儿童医院就诊的50例心脏结构异常患儿（共350张图像）的静态心脏彩色多普勒图像作为外部验证数据。依据超声工作年限将21名基层进修超声医师分为初级、中级、高级3组，比较OBICnet模型与3组医师在外部验证集中的识别效能。

与ResNet50、GBCnet模型相比，OBICnet模型的识别性能最佳，F1分数、准确率、精确率最高，分别为97.0%、98.3%、96.7%；误诊率、漏诊率最低，分别为1.6%、3.0%。在外部验证集，OBICnet模型对正常与异常图像的识别准确率、精确率、特异度分别为94.6%、91.4%、97.6%，高级、中级、初级3组医师的识别准确率分别为89.4%、83.3%、67.8%，精确率分别为78.9%、73.5%、39.3%，特异度分别为91.5%、91.6%、81.4%，OBICnet模型与3组医师比较，差异均具有统计学意义（校正P值均<0.017）；OBICnet模型对正常与异常图像的识别敏感度高于中级、初级医师组（88.5% vs 61.5%、31.8%，校正P值<0.017）；OBICnet模型漏诊率、误诊率分别为11.5%、2.4%，低于3组医师，差异具有统计学意义（校正P值均<0.017）。

OBICnet模型对正常与异常心脏超声图像的识别性能较好，在儿童先天性心脏病超声筛查工作中具有一定的应用价值。

To explore the value of artificial intelligence in ultrasound screening of congenital heart disease in children.

A total of 8543 static color Doppler ultrasound images of normal and abnormal cardiac structures from the Children's Hospital Affiliated to Zhejiang University School of Medicine from September 2021 to February 2022 were selected and divided into a training set of 6871 images, a validation set of 833 images, and a test set of 839 images at a ratio of 8:1:1. The OBICnet model was constructed, and its recognition performance was evaluated using F1 score, accuracy, precision, recall, specificity, misdiagnosis rate, and missed diagnosis rate. The performance of the OBICnet model was compared with that of the ResNet50 and GBCnet models. Additionally, 350 static color Doppler images of 50 children with abnormal cardiac structures who visited the Children's Hospital Affiliated to Zhejiang University School of Medicine from November 2022 to January 2023 were collected as external validation data. Twenty-one grassroots ultrasound physicians were divided into three groups based on their years of ultrasound experience: junior, intermediate, and senior. The recognition performance was compared between the OBICnet model and the three groups of physicians on the external validation set.

Compared with ResNet50 and GBCnet models, the OBICnet model had the best recognition performance, with the highest F1 score, accuracy, and precision, which were 97.0%, 98.3%, and 96.7%, respectively, and the lowest misdiagnosis rate and missed diagnosis rate, which were 1.6% and 3.0%, respectively. In the external validation set, the OBICnet model's recognition accuracy, precision, and specificity for normal and abnormal images were 94.6%, 91.4%, and 97.6%, respectively. The recognition accuracy of senior, intermediate, and junior physicians was 89.4%, 83.3%, and 67.8%, respectively, the precision was 78.9%, 73.5%, and 39.3%, respectively, and the specificity was 91.5%, 91.6%, and 81.4%, respectively. The differences in recognition performance between the OBICnet model and the physician groups were statistically significant (all adjusted P<0.017). The recognition sensitivity of the OBICnet model for normal and abnormal images was higher than that of the intermediate and junior physician groups (88.5% vs 61.5% and 31.8%, adjusted P<0.017). The missed diagnosis rate and misdiagnosis rate of the OBICnet model were 11.5% and 2.4%, respectively, which were significantly lower than those of the three groups of physicians (all adjusted P<0.017).

The OBICnet model exhibits superior recognition performance for normal and abnormal cardiac ultrasound images, and it has appreciated application value in the ultrasound screening of congenital heart disease in children.