Comparative evaluation of left ventricle segmentation using improved pyramid scene parsing network in echocardiography

Abstract

Automatic segmentation of the left ventricle is a challenging task due to the presence of artifacts and speckle noise in echocardiography. This paper studies the ability of a fully supervised network based on pyramid scene parsing network (PSPNet) to implement echocardiographic left ventricular segmentation. First, the lightweight MobileNetv2 was selected to replace ResNet to adjust the coding structure of the neural network, reduce the computational complexity, and integrate the pyramid scene analysis module to construct the PSPNet; secondly, introduce dilated convolution and feature fusion to propose an improved PSPNet model, and study the impact of pre-training and transfer learning on model segmentation performance; finally, the public data set challenge on endocardial three-dimensional ultrasound segmentation (CETUS) was used to train and test different backbone and initialized PSPNet models. The results demonstrate that the improved PSPNet model has strong segmentation advantages in terms of accuracy and running speed. Compared with the two classic algorithms VGG and Unet, the dice similarity coefficient (DSC) index is increased by an average of 7.6%, Hausdorff distance (HD) is reduced by 2.9%, and the mean intersection over union (mIoU) is improved by 8.8%. Additionally, the running time is greatly shortened, indicating good clinical application potential.