Abstract
Ultrasound image quality is often degraded by complex acoustic phenomena, making pixel-wise reliability estimation important for downstream analysis. Existing physical methods struggle to model complex acoustic effects, while end-to-end learning-based approaches lack interpretability and adaptability across imaging settings. We propose a de-amortized framework that first uses a neural network to predict sparse reliability cues, and then infers a dense confidence map through per-image optimization under an explicit probabilistic model. Experiments on multiple ultrasound settings show that our method achieves statistically significant improvements in expert-annotation agreement over the best baseline, with a +32.4% relative gain on the challenging fetal brain domain, and delivers the best performance across all bone shadow segmentation metrics. Our implementation will be made available upon publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_024.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=QMFRkbu1tD
BibTex
@InProceedings{GaoYuh_Learning_MICCAISAT2026,
author = { Gao, Yuhuai AND Bacher, Valentin AND Namburete, Ana I. L.},
title = { { Learning to Annotate, Optimizing to Estimate: A De-amortized Inference Framework for Ultrasound Confidence Maps } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}