Abstract
Fetal and cardiac ultrasound biometry depends on accurate landmark localization, yet a single model must handle very different anatomical views under severe label scarcity. We address nine such tasks, four cardiac and five obstetric, with one shared model trained on 5,407 labeled and 187,744 unlabeled images. The system pairs a DINOv2 ViT-S/14 encoder with lightweight per-task heatmap heads, adapted in two stages by linear probing and then fine-tuning. It learns from the abundant unlabeled scans through weak-to-strong consistency with an exponential moving average teacher and a per-task confidence gate, and decodes landmarks with distribution-aware (DARK) sub-pixel refinement averaged across five scales. On the challenge’s held-out set of 619 images it reaches a mean radial error (MRE) of 24.28 across the nine tasks; on the seven scored consistently through the challenge (two were rescored mid-way) it reaches 27.91 against the baseline’s 37.97. We further find that reliable model selection in this setting depends on subject-grouped splits and a stable evaluation protocol.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_092.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=P9ojyliHnQ
BibTex
@InProceedings{MedRod_MultiTask_MICCAISAT2026,
author = { Medellin-Robles, Rodrigo AND Juárez-Castro, Flavio A. AND Bukhari, Mohammad Abdullah},
title = { { Multi-Task Landmark Detection for Fetal and Cardiac Ultrasound Biometry under 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}
}