文献速递妇产影像儿科影像技术与安全US

混合CNN-Transformer模型用于产前脊柱裂超声识别

A Hybrid CNN-Transformer Model With Crossover Boosted Cheetah Optimization for Prenatal Spina Bifida Identification From Ultrasound Images.
PubMed 文献 · J Ultrasound Med · 2026-07-16

本研究提出一种混合卷积神经网络-Transformer模型,用于从孕中期胎儿超声图像中自动识别脊柱裂。该模型结合CNN提取局部特征与Transformer捕获全局上下文依赖,并采用交叉增强猎豹优化算法微调超参数。结果显示模型准确率为99.02%,F1-score为98.21%,推理时间0.88秒。

要点

  • 该模型针对孕中期胎儿超声图像进行脊柱裂自动识别。
  • 模型推理时间仅0.88秒,提示可用于实时产前超声筛查场景。
  • 报告准确率为99.02%、F1-score为98.21%。

核心结论

  • 提出的混合CNN-Transformer模型为产前脊柱裂自动检测提供了有效且临床可行的解决方案。

研究设计与样本

  • 研究类型为深度学习模型开发与验证研究,具体样本量与中心数原文未提供。
  • 图像来源为孕中期胎儿超声图像。
  • 模型架构为混合卷积神经网络-Transformer,集成CNN局部特征提取与Transformer全局上下文依赖捕获。
  • 采用多分辨率CNN模块与位置编码增强空间特征表示,并利用自注意力机制突出临床相关区域。
  • 超参数通过交叉增强猎豹优化算法(结合Cheetah优化器与交叉机制)进行微调。

诊断标准与关键数值

  • 模型诊断性能:准确率为99.02%。
  • 模型F1-score为98.21%。
  • 模型推理时间为0.88秒。

临床应用与阅片要点

  • 该模型面向孕中期胎儿超声图像中脊柱裂的自动识别。
  • 自注意力机制用于突出临床相关区域,可辅助聚焦可疑病变部位。
  • 推理时间较低,提示具备实时产前超声筛查应用的潜在可行性。

局限性

  • 原文未提供样本量、数据集划分方式及外部验证信息,模型泛化性尚不明确。
  • 原文未报告与医师诊断或金标准对照的敏感度、特异度及AUC值。

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: Spina bifida is a birth defect caused by the incomplete closure of the neural tube around the spinal cord. Earlier, various deep learning-based models were developed, but they failed to capture the fine details from the ultrasound images that reduce the detection accuracy. Hence, to overcome these challenges, an automated and trustworthy deep learning framework is significant for accurate identification of spina bifida from second-trimester fetal ultrasound images.

METHODS: This research proposes a novel hybrid convolutional neural network-transformer model for spina bifida identification, where convolutional neural network integration enables the model to extract local features and transformer encoder integration enables the model to capture global contextual dependencies. Further, spatial feature representation is enhanced by using the multi-resolution convolutional neural network blocks and positional encoding. On the other hand, clinically relevant regions are highlighted through the use of a self-attention mechanism within the proposed model. The hyperparameters of the proposed model are fine-tuned using a crossover boosted cheetah optimization algorithm, which integrates the strengths of the Cheetah optimizer and the Crossover Mechanism. Thereby, optimizing the performance of the proposed model and making the model robust to spina bifida identification.

RESULTS: The proposed model achieves robust diagnostic performance, which attains an accuracy of 99.02%, an F1-score of 98.21%, and a lower inference time of 0.88 second, demonstrating its suitability in real-time prenatal ultrasound screening applications.

CONCLUSIONS: The proposed model provides an efficient and clinically viable solution for automated prenatal spina bifida detection.

来源:J Ultrasound Med,查看原文。版权归原作者所有。

引用

  1. Rajasekaran, Ramesh. A Hybrid CNN-Transformer Model With Crossover Boosted Cheetah Optimization for Prenatal Spina Bifida Identification From Ultrasound Images. J Ultrasound Med. 2026. PMID: 42462242. doi:10.1002/jum.70347

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。