DiffGAN: An adversarial diffusion model with local transformer for MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic resonance imaging (MRI) is essential for diagnosis but suffers from long acquisition times and artifacts.
• Accelerated MRI techniques are vital for patient comfort and reducing motion artifacts.
• Existing Vision Transformer (ViT) and generative adversarial network (GAN) methods face challenges with high-resolution images and stable training.

Purpose of the Study:

• To develop an efficient and stable method for accelerated MRI reconstruction.
• To overcome the limitations of current ViT and GAN-based approaches in MRI.
• To improve the quality and speed of MRI image reconstruction.

Main Methods:

• Proposed a Local Vision Transformer (LVT) based adversarial Diffusion model (Diff-GAN).
• Utilized a GAN as the reverse diffusion model for large diffusion steps.
• Implemented a forward diffusion process generating Gaussian mixture noise to stabilize GAN training.
• Incorporated LVT with local self-attention for enhanced feature extraction.

Main Results:

• Diff-GAN demonstrated superior performance in accelerating MRI reconstruction.
• The method effectively captures high-quality local features and detailed information.
• Evaluated on four diverse datasets (IXI, MICCAI 2013, MRNet, FastMRI).
• Outperformed several state-of-the-art GAN-based MRI reconstruction methods.

Conclusions:

• Diff-GAN offers a promising solution for accelerated MRI reconstruction.
• The proposed LVT and adversarial diffusion approach enhance image quality and reduce acquisition time.
• This method addresses key limitations of existing techniques, paving the way for faster and more detailed MRI scans.