A review on deep learning MRI reconstruction without fully sampled k-space.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) is crucial for clinical diagnosis but limited by long acquisition times.
• Compressed sensing and parallel imaging are established MRI acceleration techniques.
• Deep learning offers novel MRI reconstruction approaches but typically requires extensive fully sampled data.

Purpose of the Study:

• To review deep learning methods for MRI reconstruction that do not require fully sampled k-space data.
• To address the challenge of applying supervised learning in scenarios lacking complete training datasets.
• To explore strategies for obtaining prior information to train reconstruction networks.

Main Methods:

• Introduction to the MRI forward model as an inverse problem.
• Discussion of the relationship between traditional iterative methods and deep learning.
• Explanation of training reconstruction networks without fully sampled data using prior information.

Main Results:

• The review highlights techniques enabling deep learning for MRI reconstruction without fully sampled data.
• It provides insights into leveraging prior information for network training.
• Potential for extending these methods to other domains lacking ground truth is discussed.

Conclusions:

• Deep learning without fully sampled data is essential for advancing MRI reconstruction.
• The reviewed methods can be adapted for applications where ground truth is unavailable.
• Combining traditional and deep learning approaches may yield superior reconstruction outcomes.