Deep learning network for fully automatic left ventricle segmentation

The segmentation of left ventricle (LV) is an essential step for evaluation of LV structure and function, which is very important for some heart disease diagnosis. This study presents a new approach to segment LV endocardium and epicardium in all slices for two specific phase in cardiac cycle based on deep learning. The proposed method contains two step: (1) localization of region of interest (ROI) within whole image and removing unrelated areas to LV, (2) segmentation and delineation of LV based on the ROI extracted in previous step. Each step of the framework employ an isolated convolutional neural network (CNN). The experimental results demonstrated that the proposed network is capable of high precision LV segmentation. The model achieved 0.90 average dice measure for endocardium and 0.91 average dice measure for epicardium in middle slices.