A multi-centre, multi-device benchmark dataset for landmark-based comprehensive fetal biometry

FetalBiometry-MultiCentre-Landmarks is a comprehensive benchmark dataset for landmark-based fetal biometry estimation from 2D ultrasound, comprising 4,513 de-identified ultrasound images from 1,904 subjects acquired at three clinical sites using seven different ultrasound devices. The dataset provides expert anatomical landmark annotations for all clinically used fetal biometric measurements: head bi-parietal diameter (BPD), occipito-frontal diameter (OFD), transverse abdominal diameter (TAD), anterior-posterior abdominal diameter (APAD), and femur length (FL).
The dataset is derived from three existing sources: the Fetal Plane (FP) dataset, the HC18 head dataset, and a dataset acquired at University College London Hospital (UCLH). Starting from standard 2D ultrasound planes, expert sonographers manually annotated anatomical landmarks following ISUOG guidelines (FP and UCL subsets), or landmarks were derived via ellipse fitting from segmentation masks (HC18 subset). For HC18, image-centric preprocessing parameters were recomputed to ensure that all landmarks lie within the model’s effective field of view, correcting an earlier landmark-space mismatch. For each image, we provide precise landmark coordinates that define the start and end points of each biometric measurement.

Each sample in FetalBiometry-MultiCentre-Landmarks consists of:

A 2D ultrasound imageLandmark coordinates for fetal head measurements (BPD, OFD) when applicableLandmark coordinates for fetal abdomen measurements (TAD, APAD) when applicableLandmark coordinates for femur measurement (FL) when applicableMetadata including: pixel-to-millimetre conversion rate, de-identified subject ID, anatomical plane, acquisition device, and subject‑disjoint train/test split assignment.The dataset emphasizes challenging but clinically relevant cases, including:
Multi-centre variability with different ultrasound devices (GE Voluson E6, S8, S10, E8, 730; Aloka)Operator-dependent variability in probe handling and image framingFetal presentation variability across gestational agesAnatomical variability in landmark positions, sizes, and orientationsCross-domain differences between manual landmark annotation (FP, UCL) and ellipse-derived landmarks (HC18)FetalBiometry-MultiCentre-Landmarks is designed to support rigorous and reproducible benchmarking of automated fetal biometry algorithms, particularly for evaluating domain shift and cross-centre generalisation in image-guided obstetric ultrasound.DeclarationsOur dataset is released under the CC BY-NC-SA 4.0 license and may only be used for non-commercial purposes.