Denoising of brain magnetic resonance images using a MDB network

The denoising of brain magnetic resonance images could be important for the medical image analysis. Many algorithms have been proposed for this task, especially the deep learning ones which show great success compared with the classical image processing algorithms. Though satisfied results it achieve, they may fail to consider the contextual and attentive features during the feature learning process, and those ones could provide essential and complementary information for the feature encoding, and the poor learning of them could hinder the model to achieve a better performance. To address this challenge, in this paper, we propose a multi-dilated block (MDB) which aims to extract more contextual and attentive features during the feature extraction stage. The whole network is based on DnCNN, and the MDB is placed in the middle stage of the network to learn the contextual and attentive representations. Moreover, for improving the similarity between the noisy image and the denoised one from feature-level, we propose a perceptual loss which is able to further boost the performance of the MDB network. To validate the effectiveness of our proposed method, we conduct extensive experiments on the brain magnetic resonance images to compare the peak signal to noise ratio and structural similarity index, and the final experimental results demonstrate that our propose method could predict a higher resolution image compared with other ones.