Abstract
Ultrasound biometry requires accurate localization of anatomical landmarks across fetal, intrapartum, and cardiac imaging protocols. The target structures range from short endpoint measurements to dense multi-landmark cardiac configurations, while ultrasound appearance varies across devices, acquisition settings, and patients. We propose a unified coarse-to-fine landmark localization framework that combines a shared DINOv3 vision transformer encoder, task-specific feature pyramid necks, context-local task adaptation, high-resolution heatmaps, and learned offset refinement. We further introduce content-aware anatomical retrieval refinement, which normalizes the ultrasound field of view and uses similar training cases as a conservative inference-time geometric prior. On the released annotation consistency evaluation covering nine task subsets, the selected DINOv3 Vision Transformer Base (ViT-B) model achieves an average Mean Radial Error (MRE) of 5.64 pixels. On public validation, the full pipeline improves over an early ViT-L task-FPN baseline from 29.88 to 23.95. These results show that pretrained transformer features, high-resolution offset refinement, and content-normalized anatomical priors provide an effective foundation for general ultrasound biometry.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_046.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=Qs6ZOk9rez
BibTex
@InProceedings{RasHam_Unified_MICCAISAT2026,
author = { Rasaee, Hamza AND Rivaz, Hassan},
title = { { Unified Coarse-to-Fine Landmark Localization for Multi-Domain 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}
}