Adversarial training and dilated convolutions for compressed sensing MRI

Abstract

Compressed Sensing Magnetic Resonance Imaging (CS-MRI) provides the possibility to accelerate the acquisition with only a small amount of k-space data. Conventional CS-MRI methods are often time-consuming due to the numerous iterative steps. Recently, deep learning has been introduced to solve CS-MRI problem. In this paper, we propose a Residual Dilated model based on Generative Adversarial Networks, titled RDGAN, for fast and accurate reconstruction. We design a modified U-Net architecture which contains dilated convolutions to aggregate multi-scale information in the MRI. Also, inspired by residual learning, we adopt a short residual connection (SRC) and a long residual connection (LRC) strategies to help features flow into deeper layers directly and stabilise the adversarial training process. The experimental results demonstrate that the proposed RDGAN model achieves the state-of-the-art performance in CS-MRI on MICCAI 2013 grand challenge dataset.