Abstract
Accurate automated measurement of ultrasound biometric parameters is hindered by scanner hardware heterogeneity across clinical sites. We present a domain-generalised framework for the FU_Biometry Challenge (MICCAI 2026) that addresses scanner shift without explicit metadata through four contributions: (1) unsupervised pseudo-scanner discovery via image statistics clustering enabling Leave-One-Pseudo-Scanner-Out (LOPSO) cross-validation; (2) a Feature-wise Linear Modulation (FiLM) Task-Attention Gate inserted between a shared DINOv2 ViT-Small backbone and nine task-specific CoordAttention heatmap decoders, adding only 25,632 parameters (0.09%) while enabling structural knowledge transfer from easy tasks (Fetal Femur, MRE 1.5 px) to hard tasks (PSAX, MRE 51.6 px); (3) a Progressive Iterative Pseudo-Labeling Schedule (PIPS) with per-task dynamic confidence thresholds starting at τ=0.70 for hard tasks and tightening to τ=0.90 across three rounds; and (4) a per-task hybrid Wing Loss and MSE formulation with domain-balanced task weighting capped at 2.0× to prevent gradient collapse on high-performing tasks. Our LOPSO evaluation reveals that standard random splits underestimate domain shift by 12–18 px MRE, confirming scanner heterogeneity as the primary bottleneck. The proposed framework achieves consistent improvement across all nine clinical domains on the FU_Biometry validation set.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_058.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=R2PZdDIiyb
BibTex
@InProceedings{MisSak_DomainGeneralised_MICCAISAT2026,
author = { Mishra, Saksham},
title = { { Domain-Generalised Ultrasound Biometry via FiLM-Conditioned Multi-Task Landmark Detection with Progressive Semi-Supervised Learning } },
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}
}