Feature Tensor Low-Rank Representation Network for Semi-Supervised Echocardiography Video Left Ventricle Segmentation.

As a non-invasive and real-time diagnostic method, echocardiography plays a crucial role in assessing heart health and diagnosing disease. Accurate segmentation of echocardiography videos is essential for understanding cardiac structure and function. However, it encounters several challenges, including the presence of speckle noise, low image contrast, and incomplete video annotations. In this study, we present a novel semi-supervised approach for left ventricular segmentation in echocardiography videos based on low-rank reconstruction of video feature tensors. Specifically, we employ a shared 2D convolutional backbone to extract deep semantic features from each frame and temporally stack them into high-order feature tensors. To exploit the high spatiotemporal correlation inherent in echocardiography, we propose utilizing the tensor singular value threshold algorithm in the transform domain to represent the video feature tensor using the key semantic feature components, eliminating feature redundancy and obtaining compact video features. Our experimental evaluations on the CAMUS dataset demonstrate that the proposed method surpasses state-of-the-art techniques, achieving the most accurate left ventricle segmentation results in echocardiography videos.Clinical relevance- In this work, a deep learning methodology to segment the left ventricle in echocardiography videos is presented. The superior performance of our proposed method on the CAMUS public dataset reveals its possibilities in clinical applications.