Abstract
Ultrasound biometry relies on accurate anatomical landmark localization, yet medical ultrasound images are often affected by speckle noise, low-contrast boundaries, and acoustic artifacts, while high-quality landmark annotations remain costly. To exploit abundant unlabeled ultrasound data and reduce the natural-to-ultrasound domain gap, we propose SonoAdapt, a foundation model adaptation framework for supervised multi-task ultrasound biometry. SonoAdapt first adapts the DINOv2 backbone using 191,170 unlabeled images and the raw images from 6,768 labeled samples, without using landmark annotations. The adapted DINOv2 branch is then combined with ViTPose and ConvNeXt branches for supervised multi-task landmark heatmap prediction. During inference, fold, backbone, and test-time augmented predictions are fused in the heatmap space, followed by DARK sub-pixel decoding. We evaluate SonoAdapt on the FU-Biometry Challenge Dataset and the London Dataset. SonoAdapt achieves an average MRE/MAE of 18.90/11.73 pixels on FU-Biometry and 16.83/10.24 pixels on the London Dataset, which are the best average MRE and MAE on both benchmarks. Ablation experiments show that ultrasound-domain self-supervised adaptation brings the largest gain, with further improvements from heatmap fusion and sub-pixel decoding. Code: https://github.com/dslhash/Foundation-Model-SonoAdapt.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_034.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_034_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=vDHxlVEWoO
BibTex
@InProceedings{DonShi_SonoAdapt_MICCAISAT2026,
author = { Dong, Shilong AND Zhang, Guilin AND Wang, Dongsheng AND Liu, Bingchuan AND Zhu, Hongbo},
title = { { SonoAdapt: Foundation Model Adaptation for Supervised Multi-Task Ultrasound Biometry } },
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}
}