CLA-U-Net: Convolutional Long-short-term-memory Attention-gated U-Net for Automatic Segmentation of the Left Ventricle in 2-D Echocardiograms

Left ventricular ejection fraction is one of the important indices to evaluate cardiac function. Manual segmentation of the left ventricle (LV) in 2-D echocardiograms is tedious and time-consuming. We proposed a deep learning method called convolutional long-short-term-memory attention-gated U-Net (CLA-U-Net) for automatic segmentation of the LV in 2-D echocardiograms. The CLA-U-Net model was trained and tested using the EchoNet-Dynamic dataset. The dataset contained 9984 annotated echocardiogram videos (training set: 7456; validation set: 1296; test set 1232). The model was also tested on a private clinical dataset of 20 echocardiogram videos. U-Net was used as the basic encoder and decoder structure, and some very useful structures were designed. In the encoding part, we incorporated a convolutional long-short-term-memory (C-LSTM) block to guide the network to capture the temporal information between frames in the videos. In addition, we replaced the skip-connection structure of the original U-Net with a channel attention mechanism, which can amplify the desired feature signals and suppress the noise. With the proposed CLA-U-Net, the LV was segmented automatically on the EchoNet-Dynamic test set, and a Dice similarity coefficient (DSC) of 0.9311 was obtained. The DSC obtained by the DeepLabV3 network was 0.9236. The hyperparameters of CLA-U-Net were only 19.9 MB, reduced by ~91.6% as compared with DeepLabV3 network. For the private clinical dataset, a DSC of 0.9192 was obtained. Our CLA-U-Net achieved a desirable LV segmentation accuracy, with a lower amount of hyperparameters. The CLA-U-Net may be used as a new lightweight deep learning method for automatic LV segmentation in 2-D echocardiograms.