Which GAN? A comparative study of generative adversarial network-based fast MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Fast magnetic resonance imaging (MRI) is essential for clinical efficiency, but k-space undersampling introduces artifacts.
• Deep learning models show promise for MRI reconstruction, yet direct comparisons are limited due to varied training and validation.

Purpose of the Study:

• To conduct a comparative study of generative adversarial network (GAN)-based models for magnetic resonance imaging (MRI) reconstruction.
• To evaluate the performance of four widely used GAN architectures: DAGAN, ReconGAN, RefineGAN, and KIGAN.

Main Methods:

• Reimplementation and benchmarking of four GAN-based MRI reconstruction frameworks.
• Training and testing on diverse MRI datasets (brain, knee, liver) with varying acceleration factors (2x, 4x, 6x) using random undersampling masks.

Main Results:

• RefineGAN achieved superior performance in MRI reconstruction compared to other evaluated GAN-based methods.
• Quantitative evaluations and qualitative visualizations confirmed RefineGAN's better accuracy and perceptual quality.

Conclusions:

• RefineGAN is a highly effective GAN-based method for accelerating MRI acquisition while maintaining high image fidelity.
• This comparative study provides valuable insights for selecting appropriate deep learning models for accelerated MRI reconstruction.