A review and experimental evaluation of deep learning methods for MRI reconstruction

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning has shown success in various applications, including accelerating magnetic resonance imaging (MRI) acquisition and reconstruction.
• Neural network-based methods are increasingly applied to nonlinear image reconstruction, inspired by computer vision and image processing techniques.
• Accelerated MRI is crucial for improving patient throughput and reducing motion artifacts.

Purpose of the Study:

• To provide a comprehensive overview of neural network-based methods for improving parallel imaging in MRI.
• To consolidate and summarize recent advancements in deep learning for accelerated MRI.
• To offer a better understanding of the rapidly evolving field of deep learning in MRI.

Main Methods:

• Review of recent developments in neural network-based approaches for parallel MRI.
• Introduction to classical k-space based reconstruction methods for parallel MRI.
• Coverage of both image domain techniques with improved regularizers and k-space based methods using neural networks for interpolation.

Main Results:

• Summarizes broad categories of deep learning methods that demonstrate good performance on public datasets.
• Highlights the successful application of deep learning techniques to nonlinear image reconstruction for accelerated MRI.
• Discusses the integration of deep learning for improved interpolation strategies in k-space.

Conclusions:

• Deep learning offers significant potential for accelerating MRI acquisition and reconstruction, particularly in parallel imaging.
• Further research is needed to address limitations and open problems in the field.
• Efforts towards open datasets and benchmarks are crucial for community advancement and reproducible research.