Abstract
Automated ultrasound biometry must span prenatal, intrapartum, and cardiac domains while tolerating extreme, task-imbalanced label scarcity. We present a single unified multi-task model for the FU-Biometry challenge that detects more than 20 clinical landmarks across nine tasks with one shared DINOv2 backbone and per-task heatmap heads. Starting from the official baseline we (i) replace argmax decoding with soft-argmax coordinates trained with a Wing loss to directly optimize radial error, and (ii) add mean-teacher semi-supervised learning that exploits the challenge’s 90% unlabeled pool. In the two-part score (Mean Radial Error, MRE, and Parameter Error, MAE), the MAE is dominated (67%) by two ellipse-circumference tasks (abdominal and head circumference): our model is accurate in-distribution but over-estimates their size on the official set. (iii) A simple, submission-verified distribution-matching size calibration corrects this, reducing Avg MAE from 35.01 to 27.59 and abdominal-circumference MAE from 131.8 to 77.8 while also improving Avg MRE. (iv) A per-task selection between the generic backbone and an ultrasound-specific one (USFM) reaches an Avg MRE of 25.10 and an Avg MAE of 25.16, the lowest among the shown teams. A recurring lesson for participants: a held-out split from the spatially homogeneous training set is a poor proxy for the official distribution, so every design choice here was tuned and verified on the validation leaderboard.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_056.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=2qX5A17CSB
BibTex
@InProceedings{ParHo_AUnified_MICCAISAT2026,
author = { Park, Ho-min AND Buralkin, Ilia AND Jochum, Michael D. AND Liu, Zhandong},
title = { { A Unified Multi-Task Model for Ultrasound Biometry Landmark Detection under Extreme Label Scarcity } },
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}
}