Abstract
In this paper, we investigate the case for the discriminative perceptual architecture in magnetic resonance imaging Rician noise removal. While the existing denoising methods based on noise estimation have achieved the superior performance under the synthetic situation, it fails to remove certain real noise and lost some texture details as the overfitting noise estimation. In order to improve the noise estimation based methods for MRI denoising, we propose to use a perceptual feature extractor to calculate the perceptual loss and the discriminator structure of WGAN to reduce the noise distribution divergence. The experiment demonstrates superiority of our proposed method in removing MRI noise and reserving good texture details.