LU-NET: A REPRODUCIBLE LAPLACIAN-UNCERTAINTY FRAMEWORK FOR FETAL HEAD ULTRASOUND SEGMENTATION

Abstract

Reliable and reproducible analysis of fetal ultrasound is essential for obstetric care, underpinning accurate biometric assessment and early detection of abnormalities. Despite advances in deep learning, current segmentation pipelines remain limited by inconsistent annotation formats, restricted dataset accessibility, and a lack of standardised evaluation. This study introduces a unified framework for fetal head segmentation that integrates dataset preparation, model training, calibrated uncertainty estimation, and out-of-distribution analysis into a single reproducible pipeline. The HC18 benchmark was standardised through ellipse-to-mask conversion and deterministic data partitioning, with the validation subset reserved for operating threshold selection. A compact U-Net with dual uncertainty heads was employed. The Laplacian component was integrated at the uncertainty fusion level, not as an image pre-filter, delivering accurate anatomical delineations and interpretable pixel-wise confidence maps. Evaluation demonstrated strong agreement with reference annotations, with the LU-Net model achieving a mean Dice of 0.9716 and mean IoU of 0.9454 on the test set, confirming high overlap accuracy, precise boundary fidelity, and well-calibrated probabilities. Uncertainty maps highlighted anatomically ambiguous regions and supported moderate discrimination of atypical frames.