SDUNet粗到细分割胎儿头围的超声框架
本研究提出SDUNet框架,基于DoubleUNet并引入协同损失,用于超声图像胎儿头围自动分割与测量。在公开HC18数据集训练并在独立外部数据集验证,分割与测量性能优异,Dice为98.60%,平均绝对误差为1.12 mm。
要点
- 胎儿头围自动测量可减轻超声医生手动测量负担,需关注分割边缘贴合度。
- SDUNet无需后处理,适用于存在散斑噪声和边界模糊的超声图像。
- 该方法在独立外部队列验证,具有较好泛化能力,可用于辅助报告。
核心结论
- SDUNet框架可有效分割胎儿头部区域并提供高精度头围测量,Dice为98.60%,平均绝对误差为1.12 mm。
研究设计与样本
- 研究类型为深度学习算法开发与验证,基于DoubleUNet框架并增强协同损失,训练集为公开HC18数据集,另设独立外部验证集。
- 应用对象为胎儿头围超声图像,不施加任何后处理。
关键影像表现/征象
- 胎儿头部超声图像存在散斑噪声和边界模糊,SDUNet仍能生成生理上合理的头部轮廓。
诊断标准与关键数值
- SDUNet分割性能:Dice系数98.60%、Jaccard指数97.24%、Hausdorff距离0.70 mm、平均对称表面距离0.17 mm、边界F1-score 99.45%。
- SDUNet测量性能:平均绝对误差1.12 mm、百分比平均绝对误差0.75%。
- 验证对象为胎儿头围超声图像,结果来自独立外部数据集及HC18数据集。
临床应用与阅片要点
- 可用于胎儿头围超声图像的自动分割与测量,减少超声医生手动测量工作量。
- 阅片时可将自动分割轮廓与手动勾画对比,重点关注边界模糊或噪声明显区域的贴合性。
- 无需后处理即可得到测量值,适合集成在超声产前检查流程中辅助报告。
局限性
- 本研究为算法性能验证,原文未报告存在其他局限。
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
OBJECTIVE: Fetal head circumference (HC) measurement is a routine and indispensable examination during pregnancy, closely associated with fetal health and maternal delivery planning. To reduce the manual workload of sonographers, deep learning has been increasingly applied to this task, with promising outcomes. In this study, we propose a coarse-to-fine deep learning framework for the automatic segmentation and measurement of fetal HC from ultrasound images.
METHODS: We propose SDUNet, a DoubleUNet-based framework enhanced with a tailored synergistic loss for fetal HC segmentation. The synergistic loss integrates multiple complementary loss terms and dynamically balances their contributions during training. This adaptive mechanism encourages physiologically plausible head shapes while reducing errors caused by speckle noise and boundary ambiguity in ultrasound images. SDUNet was trained on the publicly available HC18 data set and further validated on an independent external data set, without applying any post-processing.
RESULTS: The proposed method achieved superior segmentation and measurement performance. Specifically, it obtained a Dice coefficient of 98.60%, Jaccard index of 97.24%, Hausdorff distance of 0.70 mm, average symmetric surface distance of 0.17 mm, boundary F1-score of 99.45%, mean absolute error of 1.12 mm and percentage mean absolute error of 0.75%. The method further demonstrated strong generalization on the external data set.
CONCLUSIONS: Experimental results demonstrate that the SDUNet framework can effectively segment fetal head regions and provide highly accurate HC measurements.
引用
- Liu, Hu, Liu, et al. A Coarse-to-Fine DoubleUNet Framework with Synergistic Loss for Accurate Fetal Head Circumference Measurement. Ultrasound Med Biol. 2026. PMID: 42250993. doi:10.1016/j.ultrasmedbio.2026.05.010
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。