Bilateral lightweight network for left ventricle segmentation in two-dimensional echocardiography

Automatic segmentation of the left ventricle in echocardiography is an important step in diagnosing cardiovascular diseases. In recent years, deep learning has been widely applied in the medical field and can achieve high accuracy results in the automatic segmentation of the left ventricle. However, the existing models still face some challenges in the automatic segmentation of the left ventricle, such as the balance between segmentation accuracy and efficiency. To address this problem, this study proposes a bilateral lightweight network with diversified feature extraction and fusion (BLNet-DFEF), which can accurately and efficiently segment the left ventricle of echocardiography. The entire model consists of a backbone network, low-level detail feature and high-level semantic feature extraction module, feature fusion unit and segment head. Using diversified feature information extraction methods and feature information fusion methods is able to obtain richer feature information with lower computational complexity. The model achieves dice similarity coefficient (DSC) of 0.932 and Hausdorff distance (HD) of 3.63, and only takes 7.53 s to segment 360 images in the test set. Moreover, the accuracy in calculating left ventricle volume and ejection fraction is also higher. BLNet-DFEF is compared with three known models (BiseNetV2, FCN, UNeXt), and the best segmentation performance is achieved when BLNet-DFEF is used.