Discriminative Learning Based Dual Channel Denoising Network For Removal of Noise from MR Images

The presented manuscript proposes a dual path denoising network for removal of Rician noise from Magnetic Resonance Images (MRI). Noise is an undesired property which gets induced in the MRI or any other imaging modality during the image acquisition process. In this paper we introduced a dual path denoiser incorporated using different normalization techniques and activations in both paths. The medical images are needed to be free from spurious variations to plan the correct diagnosis of the disease. The diverse features learnt by the denoiser enables the network to produce the results which are closer to the expectations in medical imaging domain. The evaluation metrics showed the enhanced capability of the denoising network for removal of Rician noise from MR Images. The network trained once produced exceedingly good results for other datasets without retraining. The SSIM and PSNR metrics showed an improvement of about 4% and 5%.