Abstract
Accurate ultrasound biometry relies on precise anatomical landmark localization, yet building a unified model across heterogeneous ultrasound tasks remains challenging. Image appearance, acquisition protocols, and anatomical scale vary substantially across clinical domains, exacerbated by the severe scarcity of large-scale annotated datasets. To address the Generalization-oriented Ultrasound Biometry Challenge, we propose a two-stage foundation-model pipeline for generalizable multi-task landmark detection across nine diverse clinical tasks. Rather than treating each task as an isolated problem, we formulate biometry as cross-domain representation learning under limited keypoint supervision and distribution shift. First, a DINOv2 encoder is adapted on unlabeled ultrasound frames via DINO-style multi-crop self-distillation between a trainable student and an EMA teacher, capturing domain-specific sonographic textures without eroding the foundation model’s semantic priors. Second, the adapted encoder is coupled with a domain-agnostic, HRNet-style multi-resolution neck designed to reconstruct localization-relevant spatial hierarchies directly from transformer patch features. Task-specific convolutional heads then utilize soft-argmax for direct, sub-pixel coordinate regression.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_054.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_054_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=lycDMDQt2X
BibTex
@InProceedings{PanAnn_SelfSupervised_MICCAISAT2026,
author = { Panagiotakopoulou, Anna AND Barmperis, Alexandros AND Katsigiannis, Vasileios E. AND Matsopoulos, George K.},
title = { { Self-Supervised Domain Adaptation with DINOv2-HRNet for Multi-Task Ultrasound 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}
}