Abstract
Robust ultrasound biometry requires accurate anatomical landmark localization across different anatomies, scanners, and acquisition protocols. This paper presents a task-conditioned heatmap localization framework for landmark localization and biometric measurement in prenatal, intrapartum, pediatric cardiac, and adult cardiac ultrasound images. The core of the method is a task-aware architecture combined with semi-supervised learning. The framework adapts a frozen ultrasound representation foundation model (URFM) with a vision Transformer (ViT)-Huge backbone using low-rank adaptation (LoRA), fuses low-, middle-, and last-layer features using task and group conditioning, and predicts landmarks through a group-aware expert adaptor and a multi-readout heatmap head. To exploit unlabeled data, we further use semi-supervised pseudo-labeling: 5,000 images with confidence scores no lower than 0.75 are merged with the manually labeled data for training. On 619 samples from nine tasks, the framework achieves a task-averaged mean radial error (MRE) of 24.26 and reports mean absolute error (MAE) values for 29 biometric endpoints.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_085.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=T6CF5FYWmg
BibTex
@InProceedings{HuaPei_TaskConditioned_MICCAISAT2026,
author = { Huang, Pei-Qiu AND Zhao, Yu AND Liu, Jieyu},
title = { { Task-Conditioned Parameter-Efficient Adaptation for Robust Ultrasound Landmark Localization and Biometric 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}
}