Accelerated RAKI reconstruction for multi-slice cardiac cine applications

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Imaging

Background:

• Accelerated Magnetic Resonance Imaging (MRI) reconstruction is crucial for reducing lengthy cardiac scan durations.
• Deep learning adaptations of k-space-based parallel imaging (PI), such as scan-specific robust artificial neural-networks for k-space interpolation (RAKI), offer nonlinearity but increase reconstruction time.
• Scan-specific training in RAKI, while reducing database needs, impacts overall efficiency.

Purpose of the Study:

• To adapt RAKI reconstruction for cardiac cine imaging by optimizing training strategies.
• To leverage spatio-temporal redundancy for faster reconstructions.
• To ensure maintained or improved image quality in accelerated cardiac MRI.

Main Methods:

• Simplified the RAKI algorithm by removing nonlinear units and reducing convolutional layers for a GRAPPA-like reconstruction.
• Accelerated RAKI training by focusing on specific slices and cardiac phases within the cine data.
• Compared the proposed method against GRAPPA, RAKI, and rRAKI using image quality metrics (PSNR, NMSE, SSIM) and reconstruction time.

Main Results:

• The proposed method achieved image quality comparable to state-of-the-art techniques.
• Average reconstruction time was reduced by approximately 40% compared to GRAPPA, RAKI, and rRAKI.
• Quantitative metrics demonstrated the efficiency and effectiveness of the accelerated reconstruction.

Conclusions:

• The adapted RAKI method provides significantly faster cardiac MRI reconstructions with comparable image quality.
• Minor "striping" artifacts were observed, potentially linked to the k-space optimization process.
• This approach enhances the efficiency of cardiac cine MRI acquisition.