Left Ventricle Segmentation in Short-Axis Cardiac MR Images using Convolutional Neural Networks

Abstract

Segmentation of the left ventricle (LV) in cardiac magnetic resonance (CMR) images is an essential prerequisite step for diagnosis of cardiovascular diseases (CVDs). Delineation of the LV is time-consuming and skill-dependent. In this paper, we propose an automated LV segmentation in short-axis (SA) cardid to remove false positive pixels. We train and validate our method using publicly available 45 subjects from Sunnybrook cardiac dataset (SCD) and 100 subjects from left ventricular segmentation challenge (LVSC) database. Dice similarity coefficients (DSC) of the endocardium and epicardium for SCD were computed as 0.89 and 0.93 respectively. Jaccard similarity coefficient (JSC) of the LV for LVSC was calculated as 0.79 and showed that it outperforms the state-of-the-art methods. Results of evaluation with transferred weights to the different datasets presented high generalizability of our algorithm.