Fetal Heart Rate Monitor

Dual Agreement Consistency Learning With Foundation Models For

SEMI-SUPERVISED FETAL HEART ULTRASOUND SEGMENTATION AND DIAGNOSIS

Kathleen M. Curran‡◦

‡ Taighde ´Eireann – Research Ireland Centre for Research Training in Machine Learning, Ireland

Abstract

Congenital heart disease (CHD) screening from fetal echocardiography requires accurate analysis of multiple stan- dard cardiac views, yet developing reliable artificial intel- ligence models remains challenging due to limited anno- tations and variable image quality.

Figure: System Model & Simulation Flow for Fetal Heart Rate Monitor

In This Work, We Pro-

pose FM-DACL, a semi-supervised Dual Agreement Con- sistency Learning framework for the FETUS 2026 challenge on fetal heart ultrasound segmentation and diagnosis. The method combines a pretrained ultrasound foundation model (EchoCare) with a convolutional network through heteroge- neous co-training and an exponential moving average teacher to better exploit unlabeled data. Experiments on the multi- center challenge dataset show that FM-DACL achieves a Dice score of 59.66 and NSD of 42.82 using heterogeneous back- bones, demonstrating the feasibility of the proposed semi- supervised framework. These results suggest that FM-DACL provides a flexible approach for leveraging heterogeneous models in low-annotation fetal cardiac ultrasound analysis.

The code is available on https://github.com/13204942/FM-

1. Introduction

Congenital heart disease (CHD) is the most common fetal structural anomaly and a leading cause of neonatal morbidity and mortality. Prenatal screening relies on ultrasound (US) examination of multiple standard cardiac planes, including the four-chamber heart (4CH), left ventricular outflow tract (LVOT), right ventricular outflow tract (RVOT), and three- vessel-and-trachea (3VT) views, which together improve di- agnostic sensitivity compared with using the 4CH view alone [1, 2]. However, accurate interpretation of these views re- quires substantial clinical expertise and remains subject to inter-observer variability. In addition, fetal cardiac structures often present subtle anatomical patterns and variable image quality, making reliable analysis challenging . Although This work was funded by Taighde ´Eireann – Research Ireland through the Centre for Research Training in Machine Learning (18/CRT/6183).

artificial intelligence has shown promise in assisting prena- tal ultrasound analysis, progress is limited by the scarcity of large, well-annotated datasets and the high cost of expert la- beling .

The FETUS 2026 challenge addresses these limitations by providing a large-scale, multi-center fetal echocardiography dataset containing 5,000 images collected from seven hospi- tals and multiple ultrasound systems. The challenge focuses on a semi-supervised setting in which only 10% of the train- ing data are annotated with both segmentation masks and di- agnostic labels. Participants are required to develop models that simultaneously perform cardiac structure segmentation and CHD classification from a single ultrasound image, en- couraging robust and efficient AI solutions for fetal cardiac assessment.

To address the semi-supervised learning (SSL) setting of this challenge, we propose Foundation Model Dual Agree- ment Consistency Learning (FM-DACL), a unified frame- work that leverages heterogeneous models to effectively uti- lize limited labeled data and abundant unlabeled data. The proposed method combines a pretrained ultrasound founda- tion model, EchoCare, with a lightweight U-Net as a com- plementary network.

These Models Are Jointly Optimized

using supervised losses on labeled samples and dual agree- ment consistency on unlabeled data. Specifically, FM-DACL enforces prediction consistency at both local (pixel-wise seg- mentation) and global (classification-level) representations, encouraging stable segmentation outputs and consistent di- agnostic predictions. By exploiting complementary inductive biases between the foundation model and the convolutional network, the proposed approach improves robustness and generalization under limited annotation conditions.

2. Methodology

This section reviews EchoCare and U-Net. Subsequently, we explain our semi-supervised framework FM-DACL in the fol- lowing sections 2.2, 2.3, 2.4, and 2.5. The overview of our method is illustrated in Fig.1.

Arxiv:2603.18119V1 [Eess.Iv] 18 Mar 2026

Fig. 1. An overview of our FM-DACL method with EchoCare and U-Net for semi-supervised image segmentation.

2.1. Preliminaries

The proposed FM-DACL adopts a heterogeneous architec- ture composed of a foundation model and a convolutional network. Specifically, we use the pretrained EchoCare

Θ (·), Which Pro-

vides strong representation capability through large-scale pretraining on echocardiography data. EchoCare serves as a high-capacity encoder for extracting rich anatomical features from fetal ultrasound images . To complement the founda- tion model, we employ a standard U-Net architecture as

Φ(·). U-Net Provides An Efficient Convolutional

encoder–decoder structure that captures fine-grained spatial details while maintaining computational efficiency . The two networks exhibit complementary inductive biases: the foundation model offers powerful global feature representa- tions, whereas the convolutional network preserves detailed local structures. This heterogeneous design forms the basis of the proposed FM-DACL framework. During training, knowl- edge is exchanged between the two networks through dual agreement consistency, enabling effective SSL from limited labeled data while leveraging the large amount of unlabeled samples available in the dataset.

2.2. Co-Training Between Cnn And Transformer

FM-DACL adopts a cross-supervision strategy between the two heterogeneous networks: the EchoCare foundation model

Φ(·). This Mechanism Encourages The

two models to learn from each other’s predictions, allowing complementary knowledge transfer during training. Given an unlabeled input batch XUL, the two networks in-

Θ (Xul) And F 2

ϕ(XUL). Pseudo labels are then generated by converting the predicted

,

where fOH(x = c) = 1[x=c] denotes the one-hot encoding operator. During optimization, the pseudo labels produced by one network are used to supervise the other network. Specifi-

Θ , While Gradients

are not propagated between predictions and their own pseudo labels. This cross-supervision mechanism allows both net- works to iteratively refine their predictions on unlabeled data and improves the robustness of the SSL process.

The cross-supervision objective is formulated as a semi- supervised loss Lcps, composed of two reciprocal terms L1

And L2

cps. Each network is trained using pseudo labels gener- ated by the other network for both segmentation and classifi-

], I̸ = J

where i, j ∈{1, 2} and i̸ = j, indicating that each net- work is supervised by pseudo-labels generated from the other network. H(·) and D(·) denote the cross-entropy and Dice losses for the segmentation task.

B(·) Denotes The Binary

cross-entropy with logits loss used for multi-label classifica-

Φ(·)

denote their corresponding classification outputs. ˜Y and ˜C represent pseudo labels for segmentation and classification generated from the counterpart network. During inference,

Θ ) Is Used To Pro-

duce the final segmentation and classification predictions.

2.3. Interpolation Consistency Learning

FM-DACL uses an interpolation consistency learning strategy

Φ(·). This Strategy Encourages The Model To

produce consistent predictions when inputs are linearly inter- polated, thereby improving generalization and robustness. Following the mean-teacher paradigm, we maintain a

(·) Whose Parameters ¯Φ Are Updated As The

exponential moving average (EMA) of the student parameters

Ul), Where

Mix(·) denotes the mixup operation with combination ratio σ = 0.5, and i + j equals the unlabeled batch size BUL. The student network prediction for the mixed input

F 2

ϕ( ˆXUL) is encouraged to match the interpolation of the cor-

2

.

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.