Abstract
Automated ultrasound biometry can support prenatal, intrapartum, and cardiac assessment, yet many measurements remain time-consuming and operator-dependent. The MICCAI 2026 Foundation Model Challenge for Ultrasound Biometry defines a heterogeneous benchmark in which one model must localize task-specific landmarks and derive biometric parameters across multiple ultrasound domains. We present a semi-supervised multi-task landmark detection framework that combines a DINOv3-pretrained shared encoder, task-specific CBAM heatmap heads, EMA teacher-student pseudo-label learning, task-aware measurement losses, and DARK-based sub-pixel decoding. On internal validation, the proposed configuration achieved a best macro-validation mean radial error (MRE) of 18.4827, improving over the DINOv2 supervised baseline by 13.98%. Original-scale evaluation showed that DARK decoding reduced both MRE and mean absolute error. These results indicate that pretrained visual representations, semi-supervised consistency learning, and measurement-aware optimization are useful for multi-domain ultrasound biometry.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_019.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=6YsjucqzPu
BibTex
@InProceedings{CuiPin_Semisupervised_MICCAISAT2026,
author = { Cui, Ping AND Yang, Zi AND Li, Jinke AND Li, Zengxing AND Fu, Yu},
title = { { Semi-supervised DINOv3 Adaptation for Multi-domain 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}
}