3D Magnetic Resonance Image Denoising using Wasserstein Generative Adversarial Network with Residual Encoder-Decoders and Variant Loss Functions

Magnetic resonance imaging (MRI) is frequently contaminated by noise during scanning and transmission of images, this deteriorates the accuracy of quantitative measures from the data and limits disease diagnosis by doctors or a computerized system. It is common for MRI to suffer from noise commonly referred to as Rician noise because the uncorrelated Gaussian noise is present in both the real and imaginary parts of a complex K-space image with zero mean and equal standard deviation, the distribution of noise in magnitude MR images typically tends to be related to Rician distributions. To remove the Rician noise from an MRI scan, deep learning has been used in the MRI denoising method to achieve improved performance. The proposed models were inspired by the Residual Encoder-Decoder Wasserstein Generative Adversarial Network (RED-WGAN). Specifically, the generator network is residual autoencoders combined with the convolution and deconvolution operations, and the discriminator network is convolutional layers. As a result of replacing Mean Square Error (MSE) in RED-WGAN with Structurally Sensitive Loss (SSL), RED-WGAN-SSL has been proposed to overcome the loss of important structural details that occurs because of over-smoothing the edges. The RED-WGAN-SSIM model has also been developed using Structural Similarity Loss SSIM. The proposed RED-WGAN-SSL and RED-WGAN-SSIM models are formed by using the SSL, SSIM, Visual Geometry Group (VGG), and adversarial loss that are incorporated to form the new loss function. They preserved the informative details and fine image better than RED-WGAN, so our models could effectively reduce noise and suppress artifacts.

@article{Sayed2023,
  title     = {3D Magnetic Resonance Image Denoising using Wasserstein Generative Adversarial Network with Residual Encoder-Decoders and Variant Loss Functions},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {8},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Hanaa A. Sayed and Anoud A. Mahmoud and Sara S. Mohamed},
  doi       = {10.14569/IJACSA.2023.0140882},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140882}
}