Abstract
The FoundUS challenge requires a single model to localise anatomical landmarks across nine heterogeneous ultrasound tasks spanning fetal, obstetric and cardiac domains, with between two and twenty-two landmarks per image. Given a fixed multi-task heatmap architecture, which design axis actually governs accuracy? We report a controlled ablation over five alternative configurations of a smaller encoder, each of which leaves the hardest task (head circumference, HC) immovable at approximately 30 px mean radial error (MRE). A single change of encoder, from DINOv2-Small to DINOv3-Large adapted with LoRA (2.39 M of 305 M parameters trainable, 0.78%), reduces average MRE from 11.01 to 8.61 px and HC from 30.5 to 19.1 px. We additionally show that detecting and cropping split-screen acquisition artefacts in cardiac views yields order-of-magnitude gains (IVC 40.2 to 2.76 px; PSAX 27.4 to 1.70 px). Finally, we quantify a covariate shift between our held-out split and the challenge validation set, and show through a controlled distortion experiment that it is attributable to acquisition geometry unseen during training. We conclude that for unified multi-task ultrasound landmark detection, representational capacity, not loss or resolution engineering, is the binding constraint.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_083.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=KYnVSZpXSU
BibTex
@InProceedings{MooSae_Encoder_MICCAISAT2026,
author = { Moosivand, Saeed AND Wang, Ziyang},
title = { { Encoder Capacity Is the Binding Constraint in Unified 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}
}