Mikołaj Kamil Wojciech, Christensen Anders Nymark, Taksøe-Vester Caroline Amalie, Feragen Aasa, Petersen Olav Bjørn, Lin Manxi, Nielsen Mads, Svendsen Morten Bo Søndergaard, Tolsgaard Martin Grønnebæk
Department of Applied Mathematics and Computer Science, Technical University of Denmark, Kongens Lyngby, Denmark.
Copenhagen Academy for Medical Education and Simulation (CAMES), Copenhagen, Denmark.
NPJ Digit Med. 2025 May 29;8(1):318. doi: 10.1038/s41746-025-01704-0.
Ultrasound assessment of fetal size and growth is the mainstay of monitoring fetal well-being during pregnancy, as being small for gestational age (SGA) or large for gestational age (LGA) poses significant risks for both the fetus and the mother. This study aimed to enhance the prediction accuracy of abnormal fetal growth. We developed a deep learning model, trained on a dataset of 433,096 ultrasound images derived from 94,538 examinations conducted on 65,752 patients. The deep learning model performed significantly better in detecting both SGA (58% vs 70%) and LGA compared with the current clinical standard, the Hadlock formula (41% vs 55%), p < 0.001. Additionally, the model estimates were significantly less biased across all demographic and technical variables compared to the Hadlock formula. Incorporating key anatomical features such as cortical structures, liver texture, and skin thickness was likely to be responsible for the improved prediction accuracy observed.
超声评估胎儿大小和生长情况是孕期监测胎儿健康的主要手段，因为胎儿小于孕周（SGA）或大于孕周（LGA）对胎儿和母亲均构成重大风险。本研究旨在提高异常胎儿生长情况的预测准确性。我们开发了一种深度学习模型，该模型基于从65752名患者进行的94538次检查中获取的433096张超声图像数据集进行训练。与当前临床标准哈德洛克公式相比，深度学习模型在检测SGA（58%对70%）和LGA方面表现显著更好（41%对55%），p<0.001。此外，与哈德洛克公式相比，该模型的估计在所有人口统计学和技术变量上的偏差显著更小。纳入诸如皮质结构、肝脏纹理和皮肤厚度等关键解剖特征可能是观察到预测准确性提高的原因。
