基于AI的多模态超声模型预测自发性早产
该前瞻性队列研究纳入721例孕20-24周接受经阴道超声筛查的孕妇,其中44例发生自发性早产。研究利用宫颈二维超声、弹性成像图像及超声参数训练深度学习与机器学习模型,构建多模态预测模型。内部验证中神经网络模型预测自发性早产的AUC为0.85(95%CI 0.73-0.96),弹性成像评分与二维评分在多数模型中贡献显著。
要点
- 多模态模型结合宫颈二维超声、弹性成像及超声参数,对中孕期自发性早产具有较好的内部验证预测效能。
- 弹性成像评分与二维评分是多数模型中的重要预测特征,阅片时应重视宫颈弹性及二维形态信息。
- 该模型有助于识别自发性早产高危孕妇,为临床早期干预提供辅助依据。
核心结论
- 利用超声参数及宫颈二维超声和弹性成像图像构建的多模态人工智能模型,在内部验证中对中孕期自发性早产展现出较好的预测能力。
研究设计与样本
- 前瞻性队列研究,招募孕20-24周接受经阴道超声筛查的孕妇,共纳入721例,其中44例发生自发性早产。
- 参与者分为自发性早产组和足月组,训练三种深度学习模型对图像评分,并开发五种监督机器学习分类器进行预测。
关键影像表现/征象
- 基于超声图像的深度学习模型中,DenseNet121表现最佳,其AUC为0.92/0.93,并据此获得相应评分。
诊断标准与关键数值
- 多模态模型中神经网络模型在验证集中预测自发性早产的AUC为0.85(95%CI 0.73-0.96)。
- 纳入构建模型的超声参数包括弹性对比指数、硬度比、外口硬度、Ratio、二维评分和弹性成像评分。
临床应用与阅片要点
- 在中孕期经阴道超声筛查中,可关注宫颈二维超声与弹性成像图像特征,结合超声参数辅助评估自发性早产风险。
- 弹性成像评分和二维评分在多数预测模型中贡献显著,提示阅片时应重视宫颈弹性及二维形态的定量评估。
局限性
- 该研究仅进行了内部验证,缺乏外部验证队列。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVES: The primary aim of this study was to develop a multimodal artificial intelligence model for predicting spontaneous preterm birth using cervical 2-dimensional ultrasound and elastography images, as well as ultrasound parameters.
METHODS: This prospective cohort study recruited pregnant women undergoing transvaginal ultrasound screening at 20-24 weeks of gestation. Participants were divided into spontaneous preterm birth group and full-term group. Three deep learning models were trained on images to compute scores via the best-performing model. Five supervised machine learning classifiers were developed and validated to predict spontaneous preterm birth based on ultrasound parameters and deep learning scores, and feature importance analysis was employed to visualize the importance of each feature in the model.
RESULTS: A total of 721 pregnant women (44 with spontaneous preterm birth) were included. In deep learning models based on ultrasound images, DenseNet121 (AUC: 0.92/0.93) model exhibited excellent performance, and corresponding scores were obtained. Elasticity contrast index, hardness ratio, external os stiffness, Ratio, 2-dimension-score, and elastography-score were included to construct the 5 machine learning models. Among the multimodal models, the neural network model showed the best predictive performance in the validation set (AUC = 0.85; 95% CI, 0.73-0.96). Feature importance analysis revealed that both elastography-score and 2-dimension-score made significant contributions in a majority of the models.
CONCLUSIONS: The multimodal artificial intelligence models using ultrasound parameters and cervical 2-dimensional ultrasound and elastography images demonstrated promising predictive capabilities in internal validation for spontaneous preterm birth in the second trimester. This offers assistance for intervention in patients at high risk of spontaneous preterm birth.
引用
- Bai, Shen, Zhang, et al. Artificial Intelligence-Based Multimodal Ultrasound Model for Prediction of Spontaneous Preterm Birth: Development and Validation. J Ultrasound Med. 2026. PMID: 42581620. doi:10.1002/jum.70412
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。