A miniature U-net for k-space-based parallel magnetic resonance imaging reconstruction with a mixed loss function

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Deep learning in MRI typically requires extensive, site-specific training datasets.
• This limits the efficiency and applicability of current deep learning models for parallel MRI.

Purpose of the Study:

• To develop a novel, miniature U-net method for k-space-based parallel MRI.
• To enable efficient network model training using scan-specific autocalibrating signal data, reducing the need for large, pre-existing datasets.

Main Methods:

• A tailored, miniature U-net architecture was designed with fewer layers and channels.
• The network was trained using autocalibrating signal data with a combined magnitude and phase loss function.
• Performance was evaluated against Robust Artificial-Neural-Networks for k-space interpolation (RAKI) and Generalized Autocalibrating Partially Parallel Acquisitions (GRAPPA) using phantom and in vivo data.

Main Results:

• The proposed miniature U-net method effectively reduced aliasing artifacts and noise.
• Achieved an acceleration factor of four for both phantom and in vivo datasets.
• Demonstrated superior performance over RAKI and GRAPPA, with improvements in structural similarity index measure (0.02-0.05) and peak signal-to-noise ratio (PSNR) (0.1-3).

Conclusions:

• The miniature U-net offers an optimal balance between network performance and training sample requirements for k-space data reconstruction.
• This deep learning approach significantly improves image quality in parallel MRI compared to existing methods.