Abstract
We describe our solution to the MICCAI 2026 Foundation Model for Ultrasound Biometry (FU_Biometry) challenge, which requires a single unified model to localize landmarks for nine tasks spanning prenatal, intrapartum and adult-cardiac ultrasound, scored by mean radial error (MRE, 50%) and the error of the derived clinical parameter (MAE, 50%), aggregated with a rank-then-average scheme. Our system is a ConvNeXt-T encoder with a light FPN decoder and per-task differentiable soft-argmax (DSNT) heads, trained on the challenge data enriched by masked-image self-supervision on ~100k unlabelled ultrasound frames and by cross-teacher pseudo-labels for the label-scarce tasks. On top of this we contribute three components that produced all of our leaderboard gains. (i) A geometry-aware readout: the intrapartum landmarks are annotated in an 800x540 label space that differs from the stored image, and the two ellipse tasks are annotated as conjugate axis endpoints, so we back-project into the annotation space and enforce a shared-centre ellipse. (ii) A protocol calibration: probing the leaderboard with per-task scalar shape rescaling reveals a systematic ~3% oversize bias confined to the two ellipse tasks—an outer-to-outer versus outer-to-inner caliper-convention difference that is invisible on our own data—and correcting it improves the official parameter error by 5.47 px. (iii) A risk-bounded fine-tuning recipe: sharpness-aware minimisation with per-task loss weights derived from official per-task feedback, followed by weight-space interpolation back toward a verified reference, with a label-free prediction-drift criterion controlling the step. The final single model scores 26.857 px MRE and 24.861 px parameter MAE (from 27.376/30.334), runs at 32 ms/image, and is packaged as the offline Docker image required by the final phase. We additionally report the methods that did not transfer, each verified on the official leaderboard.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_118.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_118_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=g1oZx5UMcU
BibTex
@InProceedings{InoKaz_ASingle_MICCAISAT2026,
author = { Inoue, Kazuyuki AND Shigematsu, Kosuke},
title = { { A Single Unified DSNT Model with Protocol-Calibrated Readout for Generalizable Multi-Task 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}
}