A Lightweight Bilateral Left Ventricle Segmentation Method for Echocardiography

Echocardiographic images, often vulnerable to speckle and noise distortion, have a fuzzy property, presenting particular difficulties in manual segmentation. In this study, we introduced an improved bilateral segmentation network named Echo-BiSeNet, which offered a high-performance and low-computation solution. We implemented four types of convolutional neural network architectures, i.e., Echo-BiSeNet with backbones Xception and ResNet18, U-Net, BiSeNetV1 (Res18), and DeepLabV3 (Res18), that were trained, validated, and tested using 19,882 images. The experimental results indicated that Echo-BiSeNet (Res18) showed the highest DSC value and the least skewness for the left ventricle segmentation in the four-chamber view (0.9257±0.0384). This model also outperformed other conventional models regarding IoU, Recall, and Accuracy (0.8639, 0.9280, and 0.9868, respectively). In addition, the proposed Echo-BiSeNet (Res18) only has half of the number of parameters of the U-Net model, which results in a good balance between accuracy and efficiency.