Fast MRI Reconstruction: How Powerful Transformers Are?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic Resonance Imaging (MRI) is crucial for clinical diagnosis but suffers from long scanning times.
• k-space undersampling and deep learning are key strategies for accelerating MRI acquisition.
• Transformers are emerging as powerful tools for image reconstruction tasks.

Purpose of the Study:

• To investigate the efficacy of transformer architectures for fast MRI reconstruction.
• To compare novel transformer-based Generative Adversarial Network (GAN) models against existing methods.
• To evaluate the impact of GANs on preserving image quality during accelerated MRI.

Main Methods:

• Development of a Swin Transformer Generative Adversarial Network (ST-GAN) for fast MRI.
• Introduction of edge-enhanced (EES-GAN) and texture-enhanced (TES-GAN) dual-discriminator GANs based on Swin Transformers.
• Comparison of proposed models with standalone Swin Transformers and CNN-based GANs using PSNR, SSIM, and FID metrics.

Main Results:

• Transformers demonstrate strong performance in MRI reconstruction across various undersampling levels.
• GAN-based transformer models significantly enhance reconstructed image quality, particularly for undersampling rates of 30% and above.
• Proposed EES-GAN and TES-GAN models show superior performance in preserving edge and texture details.

Conclusions:

• Transformer networks, particularly when integrated with GANs, offer a viable approach for accelerating MRI scans.
• The adversarial nature of GANs is crucial for improving image quality in accelerated MRI, especially at higher undersampling factors.
• The developed Swin Transformer-based GANs provide effective solutions for fast and high-quality MRI reconstruction.