Abstract
Automatic ultrasound biometry remains sensitive to operator, scanner, and protocol variation. The MICCAI~2026 FU-Biometry challenge evaluates one landmark-regression model across nine prenatal, intrapartum, and cardiac ultrasound tasks. TABS (Transfer-Aware Backbone Selection) ranked first on the validation leaderboard, with an average Mean Radial Error (MRE) of $23.37$ px. Its main design choice is not to pick the lowest in-domain model globally: in-domain error spans only $1$ - $12$ px, while evaluation error spans $12$ - $42$ px and does not track it. TABS therefore routes each task to the encoder path that transfers best: a curated-data DINOv2 ViT-S/14 path for dense, sub-pixel localization, or a scaled DINOv3 ViT-B/16 path for tasks with stronger geometry shift. DARK decoding reduces held-out macro-MRE by $31%$ without retraining, and two protocol fixes recover mis-scored cervical-length and AOP submissions. We report the first-place recipe, the observed transfer gap, and the negative results that failed under the same shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_039.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/FoundUS_039_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=5IcejzMtsU
BibTex
@InProceedings{SukDan_TABS_MICCAISAT2026,
author = { Sukhorukov, Daniil AND Antipushina, Ekaterina AND Koush, Yury AND Makarov, Ilya},
title = { { TABS: Transfer-Aware Backbone Selection for Cross-Centre 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}
}