Abstract
Automated ultrasound biometry requires accurate landmark localization across heterogeneous anatomical views, yet existing systems are often tied to a single organ, plane, or measurement protocol and remain sensitive to scarce annotations and unreliable pseudo labels. To address these issues, we present VSTA, a DINOv2-based unified semi-supervised framework for ultrasound biometry. VSTA combines a shared pretrained encoder, task-specific heatmap heads, and differentiable coordinate decoding to support multiple landmark layouts. Then, to exploit unlabeled images conservatively, an EMA teacher fuses original and flipped predictions into pseudo landmarks weighted by heatmap confidence, a soft confidence gate, and cross-view consistency. Finally, a scarcity-aware routing strategy increases unlabeled consistency for label-scarce and high-error tasks. On the platform validation submission, VSTA achieves 25.98 average MRE and 29.83 average MAE, showing its effectiveness for robust ultrasound biometry.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_072.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=WXBY6d4ztl
BibTex
@InProceedings{QiuXuw_VSTA_MICCAISAT2026,
author = { Qiu, Xuwen AND Fang, Xin AND Tong, Lyuyang AND Du, Bo},
title = { { VSTA: A DINOv2-Based Unified Semi-Supervised Framework for Ultrasound Biometry } },
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}
}