Comparison of U-Net and nnU-Net for Fetal Head Segmentation

Abstract

Accurate measurement of fetal head circumference is a key indicator of neurodevelopmental health during pregnancy and is routinely performed from ultrasound images. Manual measurement is time-consuming and operator-dependent, motivating the development of automated fetal head segmentation methods that can support subsequent head circumference estimation. In this paper, we implement and evaluate a standard U-Net trained with three different loss functions (binary cross-entropy, soft Dice, and a weighted combination of the two) on the publicly available HC18 dataset. We compare the results against nnU-Net, a self-configuring segmentation framework that automates preprocessing, network configuration, and postprocessing. The combined loss U-Net achieved a mean Dice coefficient of 0.96 after postprocessing. nnU-Net achieved a Dice of 0.98 with a standard deviation of 0.01, which is substantially lower than the standard deviation of 0.06 observed for the best U-Net variant. The results indicate that the primary advantage of the automated pipeline lies in consistency rather than mean performance alone. These findings demonstrate the effectiveness of automated fetal head segmentation and provide a basis for reliable downstream biometric analysis.