Partial Fourier reconstruction of complex MR images using complex-valued convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Conventional partial Fourier (PF) reconstruction methods struggle with accurate phase estimation due to local phase variations.
• This limitation leads to artifacts and restricts the extent of PF reconstruction in complex MR images.

Purpose of the Study:

• To develop a complex-valued deep learning approach for improved partial Fourier (PF) reconstruction of complex MR images.
• To overcome the limitations of conventional methods in handling rapid local phase variations.

Main Methods:

• Proposed a complex-valued deep learning approach utilizing an unrolled network architecture.
• The method iteratively reconstructs PF-sampled data and enforces data consistency, incorporating both magnitude and phase information.
• Evaluated the approach on spin-echo and gradient-echo data.

Main Results:

• The deep learning method outperformed conventional POCS PF reconstruction.
• Achieved superior artifact suppression and recovery of image magnitude and phase details, even with rapid local phase changes.
• Demonstrated effective reconstruction across different orientations and anatomical regions without noise amplification, even at high PF acceleration.

Conclusions:

• The proposed deep learning method effectively reconstructs MR data at low PF fractions, producing high-fidelity magnitude and phase images.
• Offers a valuable alternative to conventional PF reconstruction for phase-sensitive 2D or 3D MRI applications.