Abstract
Clinical ultrasound biometry converts ultrasound video observations into objective anatomical measurements for clinical assessment and decision-making. Rather than treating each 2D frame independently, an ultrasound video can be formulated as a temporally coherent trajectory of partial and noisy observations of an underlying biometric world state structured by anatomical landmarks and their geometric relations. Existing ultrasound pretraining methods learn from unlabeled images or videos, but mainly model appearance priors or temporal continuity, rather than a latent state that explains observations, evolves across frames, and aligns with measurement geometry. We propose USBiomWorld, a biometry-aware world model for robust landmark-based ultrasound measurement. USBiomWorld constrains the latent biometric state with masked observation modeling to recover missing ultrasound evidence, latent transition prediction to capture state evolution across video trajectories, and geometry-conditioned anchoring to align the state with landmark-derived measurement relations. The resulting representation is robust to partial observations, predictive across frames, and geometrically readable for landmark-based measurement. The pretrained encoder is then fine-tuned with a heatmap readout head to predict landmark heatmaps, from which landmark coordinates are extracted for biometric measurement. Experiments on nine ultrasound biometry datasets show that USBiomWorld consistently outperforms representative pretraining strategies, achieving the lowest average MRE of 12.09. Further representation analyses show that the learned state is both temporally predictive and geometrically readable, supporting ultrasound biometry world modeling as a generalizable foundation for clinical measurement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_016.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=F11igyePyq
BibTex
@InProceedings{MaChe_USBiomWorld_MICCAISAT2026,
author = { Ma, Chen AND Li, Yunshu AND Wang, Ersi AND Huang, Ziyang AND Wang, Yuanyuan AND Guo, Yi},
title = { { USBiomWorld: Ultrasound Biometry World Model for Robust Landmark Measurement } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17261},
month = {pending},
page = {pending}
}