Compressed Sensing MRI Reconstruction Using Generative Adversarial Network with Enhanced Antagonism

Abstract

Compressed sensing magnetic resonance imaging (CS-MRI) reduces the time to obtain reconstructed images by under sampling the k-space data. However, solving the ill-posed problem requires an expensive computational cost. Currently, state-of-the-art compressed sensing analysis uses deep learning-based methods to reconstruct magnetic resonance images, which achieves a great success. In this paper, we propose a novel network architecture named EAGAN based on the generative adversarial networks (GAN). This proposed model enhances antagonism between the generator and the discriminator. In particular, we introduce perceptual loss to update network parameters, which makes the texture of reconstructed image reaching a better level. In addition, residual learning strategies were adopted to improve learning efficiency and reconstruction quality. Our method obtained promising experimental results on both imaging quality and reconstruction speed for 1.5T brain MRI while only 20 percent of the raw k-space data are sampled using Gaussian random sampling mask. Each image is reconstructed in about 20 ms, which is suitable for real-time processing.