Abstract
We describe a unified keypoint-regression framework built on a self-supervised vision transformer (DINOv2) backbone that jointly addresses nine heterogeneous ultrasound biometry tasks (A4C, AOP, FA, FUGC, HC, IVC, PLAX, PSAX, and fetal-femur) with a shared encoder and lightweight task-specific heatmap heads. Starting from the official baseline with a public-validation MRE of 30.11, we introduce five complementary improvements: DARK sub-pixel decoding, supervised training refinements with DSNT regularization and task-aware weighting, Mean-Teacher semi-supervised consistency over approximately 191k unlabelled images, a scale-up from ViT-S/14 to ViT-B/14, and heatmap-level ensembling with rotational test-time augmentation. The final four-model ensemble achieves a Codabench MRE of 23.97, corresponding to a 20.4% relative reduction over the baseline. Ablation results show that the largest overall gains come from decoding refinement, supervised optimization, and inference-time aggregation, while semi-supervised consistency yields the clearest benefits on the most label-scarce cardiac tasks, especially PSAX and IVC. The source code is publicly available at: https://github.com/PuppetDiary/IUGC_2026_FMUB.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_063.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=9kLivvpSQI
BibTex
@InProceedings{WuYon_AUnified_MICCAISAT2026,
author = { Wu, Yongji AND Zhang, Jiansong AND Yang, Ye AND Yang, Shilong AND Shen, Linlin},
title = { { A Unified Foundation-Model Framework with Semi-Supervised Consistency and Heatmap Ensembling for Nine-Task Ultrasound Biometry Regression } },
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}
}