Cardiovascular disease (CVD) encompasses a series of diseases involving the heart and blood vessels, and is a significant global cause of mortality. Ejection fraction (EF) is a common indicator in clinical examinations, with its elevation or reduction closely linked to common cardiovascular diseases such as heart failure, myocardial infarction, and cardiac fibrosis. This study is based on training and optimizing a model using the EchoNet-Dynamic dataset to accurately and efficiently calculate EF, in order to assist clinicians in making timely diagnoses. In this work, a new semi-supervised deep learning network, named EF-Net, is proposed, which is based on U-net, incorporating deep supervision and attention modules. By using semi-supervised learning to increase training samples for enhancing the model’s image segmentation capability. In comparison to prior methodologies, our approach presents substantial enhancements across the majority of evaluation criteria within the dataset. The research outcomes signify that our proposed technique significantly refines the precision of EF calculation via cardiac ultrasound imaging, highlighting its prospective clinical utility.