A Deep Framework Assembling Principled Modules for CS-MRI: Unrolling Perspective, Convergence Behaviors, and...

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Compressed Sensing Magnetic Resonance Imaging (CS-MRI) accelerates data acquisition but faces challenges in reconstructing high-quality images from sparse k-space data.
• Conventional CS-MRI reconstruction methods are accurate but computationally intensive.
• Existing deep learning approaches offer speed but may compromise reconstruction reliability by neglecting domain knowledge constraints.

Purpose of the Study:

• To develop a novel deep framework for CS-MRI that combines the speed of deep learning with the reliability of principled iterative reconstruction.
• To enhance the efficiency and accuracy of MR image reconstruction from undersampled k-space data.

Main Methods:

• A deep framework was proposed, integrating a learning strategy with the iterative solver of a conventional CS-MRI reconstruction energy function.
• An optimal condition checking mechanism was embedded within the framework to ensure efficient and reliable reconstruction.
• The framework was applied to complex-valued data reconstruction, parallel imaging, and reconstruction with Rician noise.

Main Results:

• The proposed deep framework demonstrated reliable convergence to optimal solutions.
• The method achieved higher efficiency and accuracy compared to state-of-the-art techniques across various scenarios.
• Experiments on benchmark and manufacturer-testing images validated the framework's performance.

Conclusions:

• The developed deep framework offers a robust and efficient solution for CS-MRI reconstruction.
• This approach successfully fuses data-driven learning with established physics-based constraints for improved medical image reconstruction.
• The method shows significant potential for accelerating MR acquisition while maintaining high image quality and reliability.