Abstract
We describe our approach to the MICCAI 2026 Foundation Model for Ultrasound Biometry challenge, which requires anatomical landmark localization across nine heterogeneous fetal and cardiac ultrasound views under a parameter budget of 22 to 30 million. Our system is intentionally simple: a DINOv2 ViT-S/14 backbone at 448 px (29.4 million parameters) with lightweight per-task heatmap heads, trained with task-uniform sampling and decoded through an exactly invertible letterbox transform. Beyond the system, our principal contribution is a controlled empirical study of which design choices improve held-out performance on this benchmark and which do not. We report four findings. First, the spatial resolution of the token grid is more important than domain-matched pretraining. Second, a widely used square-root dataset-sampling heuristic substantially degrades accuracy under an equally weighted multi-task metric, because it induces catastrophic interference through the shared backbone. Third, the strength of backbone adaptation follows a U-shaped relationship with in-distribution validation error. Fourth, test- time augmentation is orientation dependent, benefiting laterally ambiguous fetal planes while harming fixed-orientation cardiac views. Our central observation is a dissociation between leakage-free validation and out-of-distribution test performance: the configuration that is best on a carefully de-leaked validation set (23.96 px) is worse on the test set (28.81 px) than an earlier model selected without validation tuning (28.43 px). We attribute this to the distinction between validation leakage and validation-to-test distribution mismatch, and we advise against validation-based model selection on this benchmark. A per-task oracle assembly attains 24.42 px on the leaderboard; we show, however, that no leakage-free selection rule we examined realizes this margin.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_098.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_098_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=r1EgXLJmWX
BibTex
@InProceedings{AfrKhu_SonoDiNoV2_MICCAISAT2026,
author = { Afroz, Khurram AND Pandey, Aarush AND Agrawal, Monika},
title = { { SonoDiNoV2: Geometry-Aware, Pseudo-Label Guided Multi-Task Ultrasound Model } },
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}
}