Abstract
Automatic ultrasound biometry requires accurate landmark localization across heterogeneous anatomical views, but annotated data are usually limited and highly imbalanced across different tasks in real-world settings. This study presents a single heatmap-based model for all nine challenge tasks in the MICCAI 2026 Foundation Model for Ultrasound Biometry Challenge, combining a shared pretrained vision foundation encoder with task-specific landmark prediction heads. We formulate the task as a multi-stage training pipeline: supervised heatmap learning with geometric augmentation and aspect-ratio-preserving preprocessing, followed by UniMatch-style semi-supervised learning and lightweight graph neural network (GNN) coordinate refinement. The challenge development set contains 6,768 labelled images across nine landmark detection tasks, with approximately 191k additional unlabelled images used for semi-supervised training. The first stage fully-supervised model, trained with geometry transformed data augmentation reduces the Mean Radial Error(MRE) from 18.08 to 15.69 pixels. Incorporating semi-supervised learning further lowers this value to 15.38 pixels. The shared GNN refiner yields the best overall result, reducing the sample-mean MRE to 15.13 pixels. These results show that geometry-preserving preprocessing provides the main performance gain, semi-supervised learning makes additional use of large-scale unlabelled data, and coordinate-level graph refinement yields a small further improvement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_064.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=d4hF5NeIFV
BibTex
@InProceedings{YanQia_AUnified_MICCAISAT2026,
author = { Yang, Qianye AND Yang, Yingyu AND Wang, Yipei AND Hu, Yipeng AND Noble, J. Alison},
title = { { A Unified Semi-Supervised Framework for Multi-Task Ultrasound Biometry Landmark Detection } },
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}
}