Abstract
Accurate anatomical landmark localization is fundamental to ultrasound biometry, yet manual identification is time-consuming and suffers from observer variability, while automated methods face substantial heterogeneity across tasks and devices and scarce annotations alongside abundant unlabeled data. We propose a Coordinate-Aware Multi-task Semi-supervised (CAMS) framework for unified ultrasound landmark detection. CAMS adopts a shared DINOv2 encoder to learn transferable representations, and equips each task with a coordinate-aware heatmap decoder and a task-specific DSNT module for continuous coordinate regression. To exploit unlabeled data, CAMS is optimized within a Mean Teacher paradigm using an epoch-dependent confidence-masked consistency strategy. On the MICCAI 2026 FU_Biometry dataset spanning nine tasks, CAMS outperforms representative semi-supervised baselines, reducing the average Mean Radial Error from 32.63 to 24.05 and the Mean Absolute Error from 33.16 to 29.53, demonstrating strong generalization across diverse clinical scenarios and imaging devices. The source code will be released upon publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_086.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=OwfxrZChKc
BibTex
@InProceedings{LiChe_CAMS_MICCAISAT2026,
author = { Li, Chenke AND Li, Xiaoshuang AND Wang, Yuanhan AND Wu, Beining AND Qin, Feiwei AND Chen, Yifei},
title = { { CAMS: Coordinate-Aware Multi-task Semi-supervised Learning for Ultrasound Landmark Localization } },
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}
}