A hybrid GAN-PDE framework with diffusion-guided residual attention for MRI brain image denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Image Processing

Background:

• MRI brain image quality is compromised by noise from acquisition and transmission.
• Noise degrades image quality, leading to information loss, blurred edges, and distorted structures, hindering accurate diagnosis.

Purpose of the Study:

• To develop a novel hybrid image-denoising framework for MRI brain images.
• To combine Generative Adversarial Networks (GANs) with a New Weighted Anisotropic Diffusion Coefficient (NWADC) based on partial differential equations (PDEs).

Main Methods:

• A two-phase approach utilizing a custom GAN with specialized encoder, bottleneck, and decoder blocks (DG-RAB, RERDB, DG-SFB).
• A hybrid loss function balancing adversarial, content, and diffusion-guided edge regularization losses for structural fidelity and sharpness.
• A novel PDE formulation with an adaptive weighted diffusion coefficient for refinement, ensuring faster convergence and anatomical boundary preservation.

Main Results:

• The proposed method demonstrated superior performance on Figshare and BraTS2020 datasets.
• Achieved high quantitative metrics including PSNR (41.186 dB), SSIM (0.9842), NRMSE (0.0087), UQI (0.9993), and IRP (92.70%).

Conclusions:

• The hybrid approach effectively suppresses noise while retaining edges and structural integrity in MRI brain images.
• This method shows significant potential for enhancing MRI diagnostics and clinical interpretation.