Cardiac diffusion tensor imaging simulation based on deep convolutional generative adversarial network

Abstract

Myocardial fiber structure is related to heart function, the research of which is therefore of great fundamental and clinical importance. Diffusion magnetic resonance imaging (dMRI) is the most promising method for non-invasively investigating in vivo cardiac fiber structures. However, the acquisition of dMRI images for in vivo heart is difficult due to the motion influence caused by heartbeat and breathing. To deal with the acquisition problems, we propose dMRI simulation for cardiac fiber structure using deep convolutional generative adversarial networks (DCGAN). As the first attempt, we simulate dMRI images of ex vivo hearts. Diffusion-weighted (DW) images of cardiac fibers are first generated by a DCGAN network that is trained with the acquired DW images, and then the diffusion tensor images are calculated for both acquired and simulated data, from which the cardiac fiber orientations, fractional anisotropy (FA) and mean diffusivity (MD) derived from acquisitions and simulations are finally compared. The experimental results show that the simulated DW images and the corresponding fiber orientations are very close to the real ones, and that the simulated FA and MD maps have the same distribution as the real ones but with some shift in the mean values.