Abstract
Ultrasound landmark detection is a fundamental component of intelligent ultrasound biometry, playing a critical role in improving the efficiency and consistency of clinical measurements. However, substantial appearance variations, anatomical diversity, and heterogeneous spatial distributions across different anatomical structures make it difficult for conventional single-task models to achieve unified landmark localization. To address this challenge, we propose a task-aware foundation model for universal ultrasound biometry. The proposed framework employs a frozen DINOv2 backbone as a unified encoder and introduces parameter-efficient task adaptation through Visual Prompt Tuning (VPT) and shared/task-specific Low-Rank Adaptation (LoRA). Furthermore, a task-adaptive heatmap decoder and a multi-level supervision strategy are developed, where diffusion-derived structural anchors generated from unlabeled ultrasound images are exploited to provide auxiliary supervision for learning more robust anatomical representations, thereby improving landmark localization across diverse anatomical structures. Extensive experiments on the official Foundation Model for Ultrasound Biometry Challenge Hidden Validation Set demonstrate that the proposed method enables automatic estimation of more than 20 clinically relevant biometric measurements across prenatal, intrapartum, pediatric, and adult cardiac ultrasound examinations, achieving an average Mean Radial Error (MRE) of 31.14 and an average Mean Absolute Error (MAE) of 33.07. These results demonstrate that parameter-efficient task adaptation effectively unlocks the potential of vision foundation models for universal ultrasound landmark detection and provides a promising paradigm for building unified medical world models for intelligent ultrasound analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_067.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=7cadWPV2Zl
BibTex
@InProceedings{HanTai_Towards_MICCAISAT2026,
author = { Han, Taiyu AND Bi, Yuan AND Navab, Nassir},
title = { { Towards Universal Ultrasound Biometry with Task-aware Foundation Models } },
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}
}