Deep Residual Learning for Accelerated MRI Using Magnitude and Phase Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated magnetic resonance (MR) imaging using compressed sensing (CS) and parallel imaging significantly reduces scan times.
• Existing MR reconstruction algorithms often suffer from high computational costs.
• Aliasing artifacts are a common challenge in accelerated MR image acquisition.

Purpose of the Study:

• To investigate deep residual learning networks for removing aliasing artifacts in accelerated MR imaging.
• To develop a computationally efficient algorithm for MR image reconstruction.

Main Methods:

• Utilized deep residual learning networks, comprising separately trained magnitude and phase networks.
• The algorithm functions as an iterative k-space interpolation or an image domain postprocessing method based on data availability.

Main Results:

• The proposed network successfully removed strong coherent aliasing artifacts that current methods could not address.
• Achieved good reconstruction results with significantly faster computational times compared to existing CS methods.

Conclusions:

• Deep residual learning networks offer a powerful solution for accelerated MR reconstruction.
• The framework demonstrates potential for immediate generation of accurate MR images.