Texture and Structure Sensitive 3D Multi-scale Deep Neural Network for MR Images Denoising

Abstract

Image denoising is a basic problem in image processing, with the aim of approximating the original image by suppressing noise in the noised image. The major objective of noise reduction is to reduce noise in natural images while preserving original characteristics and increasing signal-to-noise ratio (SNR). This paper presents a novel approach for Denoising MRI images using a variant of Generative Adversarial Network (GAN), which is named as Texture and Structure Sensitive 3D Multi-scale Deep Neural Network (TS-3D-MDNN). A revised multi scale 3D CNN model is presented as the Generator of this GAN framework in order to preserve more information. The noise is reduced by the use of Generator and a Discriminator circuit with the help of structure and texture sensitive loss model. The experimental findings reveal that the suggested technique outperforms each method individually in terms of mean square error and peak signal-to-noise ratio. The proposed TS-3D-MDNN method achieves up to 46% of PSNR.