Deep Group-Wise Angular Translation of Cardiac Diffusion MRI in q-space via Manifold Regularized GAN

Diffusion magnetic resonance imaging (dMRI) has become an indispensable tool for non-invasive characterization of fiber structures of tissues. Clinical applicability of dMRI is often shackled by trade-off between image quality and long acquisition time. We propose a novel group-wise image translation method to improve the angular resolution of cardiac dMRI data. It consists in using a generative adversarial network (GAN) model to estimate a sequence of images from given DW images acquired in a limited number of diffusion gradient directions. We embed a supervised manifold regularized term in the GAN loss function to exploit the correlation between multiple DW images acquired in different gradient directions. Experimental results on cardiac dMRI data demonstrated that our method can significantly improve the quality of diffusion tensor imaging (DTI) reconstruction.