ABSTRACT
Left ventricle segmentation is essential in providing quantitative physiological parameters for clinical diagnosis and management of cardiovascular diseases. In this chapter, we summarize the published techniques for fully automated left ventricle segmentation. We also propose a novel automatic framework to overcome the existing limitations in deep learning-based methods. In the proposed framework, a dilated fully convolutional neural network is developed to aggregate multi-scale contextual information without losing resolution. Additionally, an adversarial network was designed to extract global features and enforce higher-order consistency between the ground truth and predicted segmentation. Based on the output from the convolutional neural network, a guided active contour model was employed to take advantage of regional characteristics on segmented subregions and refine the boundary. The proposed method achieved a Jaccard index of 0.76 on myocardial segmentation for the LV2011 validation dataset. The trained model was also evaluated using a disjoint dataset, the Sunnybrook dataset, which encompasses cases with variable heart pathologies. Our model's performance is comparable with other published techniques developed using this dataset. The results demonstrate that our method obtains accurate segmentation that generalizes to other datasets and is robust with respect to different pathologies.
