Abstract
Ultrasound biometry is often expressed as a small set of named landmarks, although the images themselves differ substantially across fetal, intrapartum, vascular, and cardiac examinations. This paper studies a compact convolutional landmark regression framework for this heterogeneous setting. The method uses a deep attention-gated U-Net to preserve local ultrasound texture and boundary detail, followed by a task-specific heatmap head. Instead of reading coordinates with a hard argmax, we introduce a heatmap moment coordinate (HMC) layer that converts each normalized response map into a continuous landmark through its spatial first moment. Gaussian heatmap supervision and an uncertainty-weighted coordinate objective are optimized jointly. Each anatomical measurement is trained as an independent model, but the architecture, loss, validation protocol, and input-size selection rule are shared. On internal validation over nine ultrasound biometry tasks, the proposed U-Net-HMC system obtains an average radial landmark error of 28.48 pixels, an average absolute coordinate error of 17.65 pixels, and a diagonal-normalized radial error of 2.70%. Ablation against a heatmap-only argmax decoder shows that the moment layer improves most tasks, especially measurements with multiple semantically ordered points. These results indicate that a carefully regularized convolutional model can serve as a practical, interpretable baseline for multi-organ ultrasound landmark measurement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_007.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=YEnu9pieUt
BibTex
@InProceedings{YanZi_AttentionGated_MICCAISAT2026,
author = { Yang, Zi AND Cuiping AND Li, Zengxing AND Ding, Xueqian AND Fu, Yu},
title = { { Attention-Gated U-Net with Heatmap Moment Coordinates for Multi-Organ 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}
}