Dual Agreement Consistency Learning with Foundation Models for Semi-Supervised Fetal Heart Ultrasound Segmentation and Diagnosis

A research team has introduced FM-DACL (Foundation Model-based Dual Agreement Consistency Learning), a novel AI framework designed to tackle the critical challenge of screening for congenital heart disease (CHD) from fetal echocardiograms. The core problem is that training reliable AI models requires vast amounts of expertly labeled ultrasound images, which are expensive and time-consuming to produce. FM-DACL addresses this by employing a semi-supervised learning strategy that effectively leverages both a small set of labeled data and a larger pool of unlabeled images. The method's innovation lies in its 'dual agreement' mechanism, which combines a pre-trained ultrasound-specific foundation model called EchoCare with a standard convolutional neural network in a co-training setup.

This heterogeneous co-training, guided by an exponential moving average teacher model, forces the two different architectures to reach a consensus on their predictions for the unlabeled data, thereby improving the overall model's robustness and accuracy. The framework was tested on the multi-center dataset from the ISBI 2026 FETUS Challenge, where it demonstrated promising results with a Dice similarity coefficient of 59.66 and a Normalized Surface Distance (NSD) of 42.82. These metrics, while indicating room for improvement, validate the feasibility of using foundation models to bootstrap performance in data-scarce medical imaging domains. The code has been made publicly available, promoting further research and validation in clinical settings.