A robust stacking ensemble of dual backbones and self-supervised pretraining for automatic fetal head circumference measurement

Ultrasound imaging is a reliable modality in prenatal care for effective estimation of biometric parameters for fetuses to support clinical decision-making. As such, accurate detection and measurement of fetal head circumference (HC) is a commonly performed task by clinicians. However, this step is often tedious and manually extracted HCs may vary from operator to operator. Despite advances in state-of-the-art methods, achieving higher accuracy remains challenging due to the lack of publicly available large-scale datasets. This work introduces a robust end-to-end pipeline that introduces three key innovations: (1) the introduction and evaluation of the OpenFetus dataset, (2) utilizing self-supervised pre-training to extract strong feature representations on an unlabeled dataset, and (3) a fusion of two deep segmentation backbones, YOLOv12-Medium and SegFormer-B3, through a stacking ensemble meta-learner. Our proposed method reaches state-of-the-art performance on the OpenFetus dataset with Dice similarity coefficient of 97.16%, mean absolute HC error of $$0.91 \pm 1.32$$ mm, and a mean Hausdorff distance of $$0.87 \pm 0.69$$ mm, outperforming all recent baselines. This demonstrates the potential of stacking-based, self-supervised deep learning approaches for robust, reproducible, and scalable fetal biometry in real-world clinical practice.