Self-Supervised Learning for Gestational Age Estimation from Low-Cost Doppler Ultrasound in Low-Resource Settings

We present a self-supervised learning framework for gestational age (GA) estimation from low-cost, one-dimensional Doppler ultrasound recordings. A segment encoder was pretrained on unlabeled Doppler data using a hybrid approach that combines Simple Siamese Networks (SimSiam) and Variance-Invariance-Covariance Regularization (VICReg), and subsequently fine-tuned on clinically accurate GA labels. The resulting model achieved a mean absolute error of $1.19$ weeks in 5-fold cross-validation and $0.87$ weeks on an external test set, an improvement of 11\% over fully supervised approaches and 29\% over transfer learning approaches. Our approach leverages abundant unlabeled Doppler data to learn generalizable fetal signal representations, enabling accurate GA estimation in low-resource settings and yielding transferable embeddings for future maternal–fetal health applications.