A Deep Learning Framework for Fetal Heart Tracking in Ultrasound Videos: Toward Enhanced Congenital Heart Defects Detection

Congenital heart defects (CHDs) affect approximately 0.8% to 1.2% of live births worldwide and rank among the leading causes of neonatal and childhood mortality. Fetal echocardiography (FE) is a noninvasive imaging technique that has been shown to detect CHDs effectively. However, variations in data arising from heterogeneous devices, variable fetal positioning, zoom factors, scanning protocols, and acquisition methods across medical centres can pose challenges for training deep learning models and hinder their generalisation. In this paper, we propose a learning-based approach for fetal heart tracking in ultrasound (US) videos and use it to address those challenges posed by real-world clinical data. Experiment results show that incorporating temporal information significantly improves tracking accuracy and achieves superior performance compared with one of the state-of-the-art object-tracking approaches (YOLO11+BoT-SORT). The proposed method is developed using 738 scans for 401 patients from John Radcliffe Hospital, and achieved Average Precision and Intersection of the Union score of 0.866 and 0.693 respectively, and further validated on a holdout test set from a different institute. To ensure the reproducibility and further development of this research, we make the code and the trained model weights publicly available as a fully open-source tool, with an interactive annotation tool for tracking fetal hearts in US videos. The source code for these tools is available at https://github.com/QianyeYang/FEHT.