Abstract
Ultrasound biometry across prenatal, intrapartum, and cardiac domains faces inter-domain visual heterogeneity, extreme annotation imbalance, and the challenge of exploiting large unlabeled clinical image pools. Existing multi-task pipelines commonly rely on task-agnostic feature adaptation, low-resolution heatmaps, and uniform consistency supervision despite substantial differences in task semantics and annotation availability. We propose UniUSB, a domain-aware adaptation framework for unified multi-domain US biometry submitted to the MICCAI 2026 Foundation Model for Ultrasound Biometry Challenge (FU_Biometry 2026). UniUSB combines three complementary components. Task-Aware Prototype Conditioning (TAPC) injects per-task residual prototypes into the shared backbone, providing dedicated adaptation channels with negligible parameter overhead. Precision-Aware Heatmap Decoding (PAHD) replaces 64$\times$64 MSE heatmaps with 128$\times$128 AWing-supervised predictions and DARK sub-pixel decoding to reduce localization quantization error. Domain-Adaptive Semi-Supervised Training (DASS) extends Mean Teacher with Dynamic Confidence Masking and reliability weighting to exploit 191,170 unlabeled images while mitigating pseudo-label noise in data-scarce tasks. Our UniUSB (Team: NPM) ranks 8th among 59 teams on the MICCAI 2026 FU_Biometry 2026 benchmark. Relative to the baseline, the final model improves MRE across all three clinical domains, while ablation studies reveal complementary benefits and domain-specific trade-offs among the proposed components.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_078.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=tBh8ZK0mPv
BibTex
@InProceedings{ZhaYix_UniUSB_MICCAISAT2026,
author = { Zhang, Yixuan AND Xu, Qing AND Li, Yue AND He, Xiangjian AND Chen, Zhen},
title = { { UniUSB: Domain-Aware Adaptation of Vision Foundation Models for Unified 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}
}