Abstract
Reliable ultrasound biometry requires accurate landmark localization, consistent endpoint identity, and robustness to acquisition variation. We present a ConvNeXt framework for the FoundUS 2026 challenge that combines stable landmark representation with adaptation from unlabeled appearance. A ConvNeXt Small encoder and separate heatmap heads predict landmarks across nine measurement tasks. Cardiac endpoint identities remain fixed within the model throughout training and fold aggregation, while a single deterministic conversion is applied only at the public output boundary. This prevents predicted coordinate order from redefining channel correspondence. During adaptation, the encoder and heatmap heads are frozen, while residual adapters learn from labeled images whose appearance statistics are transferred from unlabeled images of the same task. Landmark targets always come from manual annotations, so no pseudo coordinates are introduced. All 182,870 deduplicated unlabeled images are used in every fold under an audited protocol. In evaluation over five folds grouped by sequence, the proposed adaptation improved the combined proxy in every fold relative to a matched labeled donor control, reducing macro mean radial error by 0.0367 pixels and measurement proxy mean absolute error by 0.0153. The ensemble achieved an official validation average MRE of 25.4894 and average MAE of 26.2592. These results show that conservative appearance adaptation can improve a ConvNeXt framework that maintains stable landmark identity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_095.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_095_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=46dhqBZRpS
BibTex
@InProceedings{CheJun_Stable_MICCAISAT2026,
author = { Chen, Junxiao},
title = { { Stable Landmark Identity and Unlabeled Adaptation: A ConvNeXt Framework for FoundUS 2026 } },
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}
}