Related Experiment Videos
A hybrid GAN-PDE framework with diffusion-guided residual attention for MRI brain image denoising
Sreedhar Kollem1, Samineni Peddakrishna2, Satrughan Kumar3
1Department of ECE, SR University, Warangal, Telangana, 506371, India.
Background:
Pre-processing of images is important in MRI brain images to enhance quality after the inclusion of noise which is generated in acquisition and transmission process. This noise can lead to the loss of information, blurring of edges and distortion of structures, making an accurate diagnosis difficult.
Objective:
To address these problems, a novel hybrid image-denoising framework is proposed that combines Generative Adversarial Networks (GANs) with a New Weighted Anisotropic Diffusion Coefficient (NWADC) that is based on partial differential equations (PDEs).
Methods:
The strategy is in two phases. First, a custom-made GAN is used, which has an encoder called Diffusion-Guided Residual Attention Block (DG-RAB), bottleneck layer called Residual in Enhanced Residual Dense Block (RERDB), and a final decoder called Diffusion-Guided Skip Fusion Block (DG-SFB). This design is able to improve feature learning, increase the preservation of edges and reduce noise. A hybrid loss function adaptively balances the adversarial loss, content loss, and diffusion-guided edge regularization loss to achieve structural fidelity and perceptual sharpness in training. A BN-SN Hybrid Discriminator is used in parallel to stabilize adversarial learning, enhance convergence and robustness by using both batch normalization and spectral normalization. The denoised outputs are then refined in the second phase using a novel PDE formulation with an adaptive weighted diffusion coefficient that speeds up the convergence, avoids staircase artifacts, and maintains the delicate anatomical boundaries.
Results:
The experimental results of the Figshare and BraTS2020 datasets showed that the proposed method outperforms the others with PSNR of 41.186 dB, SSIM of 0.9842, NRMSE of 0.0087, UQI of 0.9993, and IRP of 92.70%.
Conclusion:
This proposed approach was effective to significantly enhance noise suppression, edge retention, and structural fidelity, thus showing its potential for developing MRI diagnostics and clinical interpretation.
