Deep Magnetic Resonance Image Reconstruction: Inverse Problems Meet Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Undersampled k-space data is crucial for accelerating Magnetic Resonance Imaging (MRI).
• Deep learning (DL) shows significant potential for enhancing MRI reconstruction speed and efficiency.
• Traditional MRI reconstruction methods face limitations in speed and data requirements.

Purpose of the Study:

• To provide a comprehensive overview of deep learning-based image reconstruction methods in MRI.
• To categorize and explain the primary architectures of DL approaches for MRI.
• To discuss signal processing considerations for optimizing DL in fast MRI.

Main Methods:

• Review of two main categories of deep learning approaches: unrolled algorithms and non-unrolled methods.
• Explanation of the core structures and operational principles of each DL category.
• Analysis of signal processing challenges and opportunities in DL-based MRI reconstruction.

Main Results:

• Deep learning methods offer a promising avenue for accelerating MRI acquisition and reconstruction.
• Unrolled and non-unrolled DL architectures provide distinct strategies for image reconstruction.
• Identifying signal processing issues is key to unlocking the full potential of DL for fast MRI.

Conclusions:

• Deep learning significantly advances fast MRI by enabling reconstruction from limited data.
• Understanding the nuances of DL architectures is vital for developing improved MRI techniques.
• Further research into signal processing aspects will enhance the theoretical understanding and practical application of DL in MRI.