DACL framework boosts fetal ultrasound segmentation with 2.77% Dice gain

Researchers from University College Dublin and other institutions have developed DACL (Dual Agreement Consistency Learning), a novel semi-supervised framework for fetal ultrasound image segmentation. Published on arXiv and accepted to MICCAI 2026, DACL addresses the scarcity of pixel-level annotations by jointly training a lightweight convolutional network (1.47M parameters) and a Transformer-based network. It extends cross pseudo supervision (CPS) with a dual-agreement consistency loss that aligns both pixel-wise probability distributions and entropy-based confidence, suppressing unreliable pseudo-labels and enabling stable learning under extreme annotation scarcity. Additionally, mixup-based interpolation on unlabeled samples enhances robustness.

Tested on fetal head and abdomen ultrasound datasets with only 5% labeled data, DACL outperforms recent semi-supervised methods by up to 2.77% in Dice score and reduces the 95th percentile Hausdorff distance (HD95) by up to 14.69 mm, indicating significantly better boundary accuracy. The framework's deployment-oriented lightweight CNN makes it practical for clinical settings. Code is available on GitHub. This work demonstrates that agreement-based consistency learning can dramatically reduce the annotation burden for automated fetal ultrasound analysis.