Using Simulated Data to Predict Birthweight from Prenatal Ultrasound Images

Fetal weight estimation during pregnancy is critical for monitoring growth and planning delivery, yet current ultrasound-based estimations are prone to error. This study leveraged open access ultrasound images and simulated clinical data to develop a deep learning pipeline for fetal weight prediction. The dataset consisted of n=1,570 ultrasound images from 333 pregnant individuals at 18-40 weeks’ gestation. Relevant maternal and fetal attributes like body mass index, gestational age, amniotic fluid index, and newborn birth weight were simulated based on population averages. A convolutional neural network (CNN) using transfer learning was applied to extract imaging features. The pre-trained Xception model was adapted by adding a global average pooling layer and freezing layers during feature extraction. Extracted imaging features were combined with clinical metadata and input into linear regression and random forest models. Hyperparameter tuning via randomized search identified a random forest configuration minimizing error. On a 20% test set, the pipeline achieved a mean absolute error of 328.6 grams in predicting birth weight, suggesting feasibility of the approach. While clinical conclusions are limited without real-world data, this study demonstrates a pipeline leveraging open science resources to address an important maternal-fetal health problem. Tailored CNN architectures and further model optimization on population-representative data could enhance performance and clinical utility. This work highlights the potential of open datasets to enable rapid scientific progress and proposes a deep learning methodology for improved fetal weight assessment during pregnancy.