Abstract
Landmark-based ultrasound biometry reduces a broad spectrum of clinical measurements to a small set of anatomically defined points, yet the resulting learning problem is highly heterogeneous: the same coordinate representation is used for fetal, intrapartum, vascular, and echocardiographic structures whose image appearance, spatial scale, and point semantics differ substantially. This work investigates a high-resolution DINOv3-DSNT formulation for multi-organ ultrasound landmark regression. A DINOv3 ViT-base encoder provides dense self-supervised visual features, while a compact multi-scale decoder produces task-specific landmark heatmaps. Coordinates are obtained through a differentiable spatial-to-numerical transform, allowing heatmap supervision and continuous coordinate regression to be optimized in a single end-to-end model. Rather than imposing one shared landmark vocabulary across unrelated anatomical measurements, each task retains its own output head, native-resolution training scale, and checkpoint. Internal validation across nine ultrasound biometry tasks yields an average mean radial error of 22.83 pixels, an average mean absolute coordinate error of 14.03 pixels, and an average diagonal-normalized radial error of 2.20% in the original image coordinate system. The results indicate that dense self-supervised representations, when coupled with differentiable heatmap regression, provide a coherent technical route for high-resolution ultrasound biometric localization across diverse organs and measurement geometries.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_044.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=SQWpVIEVFo
BibTex
@InProceedings{HouJio_HighResolution_MICCAISAT2026,
author = { Hou, Jiongpei AND Li, Meng AND Xing, Yushan AND Peng, Hong},
title = { { High-Resolution DINOv3-DSNT Landmark Regression for Multi-Organ 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}
}