Enhanced Partial Fourier MRI With Zero-Shot Deep Untrained Priors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Partial Fourier (PF) acquisition accelerates Magnetic Resonance Imaging (MRI) by undersampling k-space data.
• Reconstruction of undersampled MRI data is crucial for maintaining diagnostic image quality.
• Traditional methods often struggle with artifacts or require extensive training data for deep learning approaches.

Purpose of the Study:

• To develop a novel zero-shot unsupervised deep learning method for robust partial Fourier MRI reconstruction.
• To integrate a phase constraint with a generative prior-based deep learning framework for improved reconstruction.
• To evaluate the method's performance across diverse datasets, including multi-contrast and low-field MRI.

Main Methods:

• Utilized a zero-shot deep learning framework based on untrained generative priors for MRI reconstruction.
• Integrated the virtual conjugate coil (VCC) phase constraint into the zero-shot learning approach.
• Applied the method to the fastMRI, QALAS multi-contrast, and a low-field MRI dataset for validation.

Main Results:

• The proposed zero-shot method significantly improved reconstruction quality compared to existing techniques.
• Achieved substantial reductions in Normalized Root Mean Square Error (NRMSE), ranging from 1.15x to 2.5x.
• Demonstrated robust performance across different datasets, highlighting its generalizability.

Conclusions:

• The novel zero-shot unsupervised deep learning approach offers a powerful tool for partial Fourier MRI reconstruction.
• Eliminates the need for training data, making it highly valuable for challenging data acquisition scenarios.
• Provides a robust and efficient method for enhancing MRI reconstruction quality and accelerating imaging.