Perceptually Motivated Generative Model for Magnetic Resonance Image Denoising

Area of Science:

• Biomedical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) data is often corrupted by mixed Gaussian-impulse noise.
• This noise degrades image quality and hinders downstream analysis in biomedical applications.

Purpose of the Study:

• To propose a novel deep generative model, GenMRIDenoiser, for effective noise removal in MRI.
• To address the challenges of mixed noise scenarios in biomedical image processing.

Main Methods:

• Utilized Wasserstein Generative Adversarial Network (WGAN) to overcome GAN training issues.
• Incorporated a perceptually motivated loss function to preserve image details.
• Employed batch renormalization for improved performance with non-iid data.
• Integrated a global feature attention module (GFAM) to capture long-range dependencies.

Main Results:

• The GenMRIDenoiser model demonstrated effective removal of mixed Gaussian-impulse noise from MRI data.
• Experimental results on synthetic and real MRI scans confirmed the model's utility across various degradation levels.
• The proposed method successfully preserved high-frequency components and low-level image details.

Conclusions:

• GenMRIDenoiser offers a robust solution for denoising corrupted MRI scans.
• The model shows significant potential for improving the quality and utility of biomedical images.
• This deep learning approach advances the field of low-level vision for medical imaging.