Abstract 16899: Fully-Automated Left Ventricle Segmentation Using a Dilated and Adversarial Deep Learning Architecture

Abstract

Introduction: Cardiovascular magnetic resonance (CMR) imaging is currently the gold-standard to analyze cardiac morphology and evaluate global and regional left ventricle (LV) function. Quantitative parameters such as ejection fraction (EF) and LV mass are important indicators for diagnosis and associated with morbidity and mortality. In practice, the manual or semi-automatic delineations of LV contours for function evaluation are time-consuming and prone to intra- and inter-observer variability. The objective of this study is to develop an accurate and fully-automatic LV segmentation method for CMR slices in short axis using deep learning techniques. Methods: In this study, we trained a model using the LV2011 training dataset, which contains 100 cases of patients with myocardial infarction from multiple MR scanner system using balanced steady-state free precession. We developed a novel fully convolutional network (FCN) with dilated convolutions for pixel-wise prediction. An adversarial network is added to enforce high-order consistency. Furthermore, a guided active contour model is deployed as a post-processing method that considers the intensity homogeneities and boundary smoothness of segmented components. The similarity between the automatic segmentation from our model and the manual segmentation are evaluated using Jaccard index and Hausdorff distance (HD). Results: On the LV2011 test dataset, our model achieved a Jaccard index of 0.77 and HD of 4.12 mm, outperforming existing fully-automatic methods. Examples of segmentation results are shown in Figure 1. The clinical parameter estimations also have high reliability. The correlations between the end-diastolic volume, end-systolic volume, EF, and LV mass, estimated from automatic segmentation and that from manual segmentation on the test dataset, are 0.99, 0.99, 0.93, and 0.96, respectively. Conclusions: Our results demonstrate that the proposed method obtains accurate segmentation and is robust over cases of high variability.