3D MRI Cardiac Segmentation Under Respiratory Motion Artifacts

Abstract

In cardiac magnetic resonance (CMR) imaging, 3D segmentation of the heart is important for detailed description of its anatomy and thus estimation of clinical parameters. However, images acquired in the clinical routine often contain artifacts caused by respiratory motion due to acquisition time and respiratory/cardiac motion limitations. Segmentation of these low-quality images using conventional methods often does not yield accurate results. Here, we propose a DenseBiasNet combined with Variational Auto-Encoder (VAE) network framework, a segmentation model robust to respiratory motion artifacts, for automatic CMR image segmentation from extreme images. Given an image with respiratory motion artifacts as input, DenseBiasNet is utilized as the primary branch for segmentation, and VAE network is utilized as the secondary branch to map the low-resolution image of the DenseBiasNet encoded portion as input to the low-dimensional space for reconstruction.