Automated segmentation of left ventricular myocardium using cascading convolutional neural networks based on echocardiography

Abstract

Quickly and accurately segmenting the left ventricular (LV) myocardium from ultrasound images and measuring the thickness of the interventricular septum and LV wall play an important role in hypertrophic cardiomyopathy. However, the segmentation of the LV myocardium is a challenging task due to image blurring and individual differences. We attempted to perform LV segmentation in ultrasound images using the encoder-decoder architecture of U-Net and other networks and found it to be not accurate enough. Therefore, we propose a novel multi-task cascaded convolutional neural network (called MTC-Net) to segment the LV myocardium from echocardiography. MTC-Net contains two parts: One is pre-trained Resnet-34 followed by two decoder branches for mask and boundary detection, and the other module is pre-trained with many improved novel encoder-decoder architectures for extracting more detailed features. Both parts of the network use the atrous spatial pyramid pooling module to capture high-level text information. A hybrid loss function is engaged for mask and contour prediction. The network is trained and evaluated with echocardiographic images, which are labeled manually by doctors. The comparison study with other networks shows that MTC-Net has better accuracy and performance. MTC-Net achieves state-of-the-art performance on the test set. The mean value of the dice coefficient is 0.9442 and the mean value of intersection over union is 0.8951.