Abstract

Ultrasound biometry combines landmark localization across anatomically and statistically heterogeneous tasks. The FoundUS challenge includes nine prenatal, intrapartum, and cardiac ultrasound tasks whose training-set sizes, landmark definitions, and derived measurements differ substantially. We propose a unified multi-task system using fully fine-tuned DINOv2 encoders, task-specific localization heads, and balanced task sampling. To address task-dependent errors, we construct a hierarchical ensemble combining DSNT, residual log-likelihood estimation, SimCC, and a dual RLE–SimCC head, together with diversity in training partitions, initialization, encoder capacity, and input resolution. The final system contains 17 checkpoints organized into three equally weighted blocks. On the official validation server, it achieved an Average mean radial error of 24.233 and an Average parameter mean absolute error of 29.345. Relative to the MRE-optimal two-block ensemble, the selected system changed Average MRE by only 0.015 pixels while improving parameter error on seven of nine tasks. These results support structured ensemble diversity for unified ultrasound landmark localization under substantial task heterogeneity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_102.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=dy1YqoOdCG

BibTex

@InProceedings{SemIli_Hierarchical_MICCAISAT2026,
        author = { Semenkov, Ilia},
        title = { { Hierarchical Ensembles of Complementary Landmark Detectors for 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}
}


back to top