Automatic segmentation of left ventricle using parallel end-end deep convolutional neural networks framework

Abstract

Abstract Under the background of high incidence and mortality of cardiovascular diseases, the accurate and automatic left ventricle (LV) segmentation method is of essential importance for the diagnosis of the cardiovascular system. However, fully automatic LV segmentation remains challenging due to the complex structure of cardiac magnetic resonance image (MRI) and the morphological changes of LV caused by various cardiovascular diseases. In this paper, we propose a novel parallel end-to-end convolutional neural network (CNN) for LV segmentation. Our network consists of two interactive subnetworks which utilize essentially identical but formally different labels in the hope that they can learn segmentation from different perspectives. The two subnetworks take the same cardiac MRI as input and output a pair of segmentation maps in different forms. After averaging the two segmentation maps obtained from the two subnetworks, we get the final contours of the endocardium (endo) and epicardium (epi) simultaneously. The proposed method is evaluated on the dataset provided by the Left Ventricle Full Quantification Challenge of MICCAI 2019. The average Dice scores on epi, endo, and myocardium (myo) reach 0.961, 0.949, and 0.867 respectively which outperform the other methods. The experimental results show that our method has the potential for clinical application.