Super-resolution of magnetic resonance images using Generative Adversarial Networks

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) faces limitations including small spatial coverage, high costs, and long scan times.
• Accelerating MRI acquisition by reducing measurements is crucial for overcoming these limitations.
• Machine Learning (ML) techniques, particularly super-resolution (SR), show promise in recovering high-resolution (HR) images from low-resolution (LR) signals.

Purpose of the Study:

• To review Generative Adversarial Network (GAN)-based super-resolution (SR) methods for Magnetic Resonance Imaging (MRI).
• To evaluate the capability of GANs in upscaling MRI data by a factor of ×4 while preserving critical details.
• To assess the potential of GAN-based methods in reducing medical costs and improving patient experience.

Main Methods:

• Review of existing literature on GAN-based SR techniques applied to MRI reconstruction.
• Analysis of quantitative and qualitative performance metrics for different GAN models.
• Comparison of GAN-based SR methods against other deep learning approaches for MRI upscaling.

Main Results:

• GANs demonstrate significant potential in MRI reconstruction and acceleration.
• SRResCycGAN quantitatively outperforms other deep learning methods in recovering ×4 downscaled MRI images.
• Beby-GAN achieves superior perceptual quality in qualitative assessments, highlighting GANs' ability to infer missing details.

Conclusions:

• GAN-based SR methods can effectively upscale MRI data by ×4, maintaining trustworthy high-frequency details.
• These methods offer a pathway to reduce medical costs, decrease patient distress, and enable novel MRI applications.
• GANs present a powerful tool for advancing MRI technology, making it more accessible and efficient.