Generative priors for MRI reconstruction trained from magnitude-only images using phase augmentation

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Generative image priors enhance image reconstruction quality.
• Current methods often struggle with limited data or phase information in Magnetic Resonance Imaging (MRI).

Purpose of the Study:

• To develop a workflow for creating robust generative image priors from magnitude-only MRI data.
• To evaluate the impact of phase information and dataset size on prior performance.
• To compare generative priors against traditional regularization methods.

Main Methods:

• Constructing training datasets from magnitude-only MR images.
• Augmenting datasets with phase information to train complex image priors.
• Evaluating priors using linear and nonlinear reconstruction for compressed sensing parallel imaging.

Main Results:

• Priors trained on complex images (magnitude + phase) outperformed those trained on magnitude-only images.
• Larger training datasets led to more robust priors.
• Generative priors demonstrated superiority over [Formula: see text]-wavelet regularization in high undersampling scenarios.

Conclusions:

• Phase information is crucial for improving generative prior performance in MRI reconstruction.
• Large datasets enhance the robustness of generative priors.
• Generative priors offer a powerful alternative to traditional regularization for accelerated MRI.