Automatic Segmentation of Left Myocardium in CMR Based on Fully Convolutional Networks

Abstract

Myocardial segmentation plays an important role for quantitative evaluation of heart diseases and cardiac image processing and analysis. However, myocardial segmentation has always been a challenging task because gray scale intensities of the myocardium and tissues around the heart are very close and that significant differences exist in myocardial structure between different slices or slices at different times. Traditional segmentation algorithms are difficult to obtain accurate and robust segmentation results and are usually semi-automatic which require manual operations and extra workload. Therefore, the development of a fully automatic myocardial segmentation algorithm is an appealing research goal. In this paper, we propose an automatic myocardial segmentation algorithm based on fully convolutional neural networks. By building an end-to-end model, the segmentation speed has been improved without affecting the segmentation accuracy. Performance comparisons between the proposed HeartNet and state-of-art methods demonstrated the effectiveness of our algorithm, which achieved an average DSC of 90.48% by segmenting 144.9 frames per second.