Automatic segmentation of left ventricular myocardium by deep convolutional and de-convolutional neural networks

Abstract

Deep learning has been integrated into several existing left ventricle (LV) endocardium segmentation methods to yield impressive accuracy improvements. However, challenges remain for segmentation of LV epicardium due to its fuzzier appearance and complications from the right ventricular insertion points. Segmenting the myocardium collectively (i.e., endocardium and epicardium together) confers the potential for better segmentation results. In this work, we develop a computational platform based on deep learning to segment the whole LV myocardium simultaneously from a cardiac magnetic resonance (CMR) image. The deep convolutional network is constructed using Caffe platform, which consists of 6 convolutional layers, 2 pooling layers, and 1 de-convolutional layer. A preliminary result with Dice metric of 0.75±0.04 is reported on York MR dataset. While in its current form, our proposed one-step deep learning method cannot compete with state-of-art myocardium segmentation methods, it delivers promising first pass segmentation results.