文献速递产科母胎医学

基于人工智能的超声筛查用于产前检测胎盘植入谱系

Artificial intelligence-based ultrasound screening for antenatal detection of placenta accreta spectrum.
PubMed 文献 · Am J Obstet Gynecol MFM · 2026-07-31

本研究为单中心回顾性研究,纳入113例胎盘植入谱系(PAS)风险患者2018至2025年间的756个胎盘超声DICOM文件,提取38,907帧图像,构建以EfficientNetB0为骨干的卷积神经网络并结合临床风险因素的集成模型。模型预测PAS的准确率为88%(95%CI 63.6%-98.5%),敏感度100%(95%CI 66.4%-100.0%),特异度75%(95%CI 34.9%-96.8%),受试者工作特征曲线下面积0.972(95%CI 0.875-1.000),测试队列无假阴性。作者认为该模型有望作为产前PAS筛查工具,但需前瞻性试验验证。

要点

  • 集成模型预测PAS准确率88%,敏感度100%,特异度75%,AUC 0.972。
  • 测试队列无假阴性,阳性预测值81.8%,阴性预测值100%。
  • 模型变量重要性集中在胎盘界面,支持生物学合理性。

核心结论

  • 新型人工智能模型可在分娩前准确、敏感地预测胎盘植入谱系,支持其作为早期诊断筛查工具的潜在用途,尚需前瞻性试验验证。

研究设计与样本

  • 单中心回顾性研究,纳入113例PAS风险患者2018至2025年的756个胎盘超声DICOM文件,提取38,907帧灰度图像。
  • 超声时平均孕周30.89±3.67周,患者按79例训练、17例验证、17例测试分组。
  • 采用ImageNet预训练EfficientNetB0骨干网络,帧级概率平均为患者级共识评分,再结合既往剖宫产次数与前置胎盘状态训练逻辑回归、随机森林和梯度提升分类器并集成。

主要发现

  • 模型预测PAS存在与否的准确率为88%(95%CI 63.6%-98.5%)。
  • 敏感度100%(95%CI 66.4%-100.0%),特异度75%(95%CI 34.9%-96.8%)。
  • 阳性预测值81.8%(95%CI 48.2%-97.7%),阴性预测值100%(95%CI 54.1%-100.0%)。
  • 测试队列无假阴性,AUC为0.972(95%CI 0.875-1.000)。
  • 模型变量重要性评分集中于胎盘界面。

临床怎么用

  • 该模型有潜力作为产前PAS筛查工具,辅助识别高风险患者以支持更早期诊断。
  • 当前证据来自单中心回顾性小样本测试队列,临床应用前需前瞻性试验验证。

局限性

  • 单中心回顾性设计,测试队列仅17例,置信区间较宽。
  • 样本来自PAS风险患者,外推至一般人群需谨慎。

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)。

原文

BACKGROUND: Placenta accreta spectrum is a leading cause of maternal morbidity and mortality and is increasing in incidence; however, only 30% to 50% of cases are diagnosed antenatally. Although ultrasonographic findings may inform the risk of placenta accreta spectrum, multiple factors may lead to inconclusive results or misdiagnosis. OBJECTIVE: We hypothesized that a novel artificial intelligence model could be used as a screening tool to accurately predict placenta accreta spectrum through image classification of 2-dimensional placental ultrasounds. STUDY DESIGN: This single-center retrospective study included 756 placental ultrasound Digital Imaging and Communications in Medicine (DICOM) files, from which 38,907 grayscale portable network graphics frames were extracted from 113 patients at risk for placenta accreta spectrum from 2018 to 2025. The mean gestational age at ultrasound was 30.89±3.67 weeks. Patients were stratified to into training (n=79), validation (n=17), and test (n=17) groups. Images were classified by final pathologic grades. We used an ImageNet pretrained EfficientNetB0 backbone, followed by global average pooling and a regularized fully connected layer with sigmoid activation for binary classification. Frame-level probabilities from the convolutional neural network output were averaged to patient-level consensus scores. These scores, together with the number of previous cesarean deliveries and previa status, were input as variables into training a logistic regression, random forest, and gradient boosting classifier, which were ensembled for the final prediction of placenta accreta spectrum incorporating patient-identifiable risk factors. RESULTS: The convolutional neural network ensembled model predicted the presence or absence of placenta accreta spectrum accurately in 88% (95% confidence interval, 63.6%-98.5%) of cases, with a sensitivity of 100% (95% confidence interval, 66.4%-100.0%) and a specificity of 75% (95% confidence interval, 34.9%-96.8%). The positive predictive value was 81.8% (95% confidence interval, 48.2%-97.7%), and the negative predictive value was 100% (95% confidence interval, 54.1%-100.0%). There were no false negatives in the testing cohort. The area under the receiver operating characteristic curve was 0.972 (95% confidence interval, 0.875-1.000). Model variable importance scores were concentrated at the placental interface, highlighting biological plausibility. CONCLUSION: This novel artificial intelligence model achieved accurate and sensitive placenta accreta spectrum prediction before delivery. The results of the model support its potential use as a screening tool for the earlier diagnosis of placenta accreta spectrum, warranting future prospective trials. El resumen está disponible en Español al final del artículo.

来源:Am J Obstet Gynecol MFM,查看原文。版权归原作者所有。

引用

  1. Hammerquist AL, Lombaard HA, Sarada S, et al. Artificial intelligence-based ultrasound screening for antenatal detection of placenta accreta spectrum. Am J Obstet Gynecol MFM. 2026. PMID: 42537905. doi:10.1016/j.ajogmf.2026.102079

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。