Deep Recurrent Partial Fourier Reconstruction in Diffusion MRI

Fasil Gadjimuradov, Thomas Benkert, Marcel Dominik Nickel, Andreas Maier

Abstract: Partial Fourier (PF) acquisition schemes are often employed to increase the inherently low SNR in diffusion-weighted (DW) images. The resulting ill-posed reconstruction problem can be tackled by an iterative Projection Onto Convex Sets (POCS). By relaxing the data constraint and replacing the heuristically chosen regularization by learned convolutional filters, we arrive at an unrolled recurrent network architecture which circumvents weaknesses of the conventional POCS. Further, knowledge on the pixel-wise noise level of MR images is incorporated into data consistency operations within the reconstruction network. We are able to demonstrate on DW images of the pelvis that the proposed model quantitatively and qualitatively outperforms conventional methods as well as a U-Net representing a direct image-to-image mapping.