Composite MR image reconstruction and unaliasing for general trajectories using neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Rapid parallel magnetic resonance imaging (MRI) presents significant image reconstruction challenges.
• Aliasing artifacts are a primary concern in undersampled MRI data acquisition.
• Existing reconstruction methods often struggle with general sampling trajectories.

Purpose of the Study:

• To introduce a novel neural network-based image reconstruction technique for rapid parallel MRI.
• To develop a method capable of handling arbitrary data acquisition trajectories.
• To improve the quality and efficiency of MRI image reconstruction.

Main Methods:

• Proposed Composite Reconstruction And Unaliasing using Neural Networks (CRAUNN) framework.
• Utilized neural networks to learn the transformation from aliased to alias-free images.
• Operated in the image domain, independent of coil sensitivity estimation and sampling trajectory.
• Trained on densely sampled low-frequency data to generalize to sparsely sampled high-frequency data.

Main Results:

• CRAUNN successfully reconstructs alias-free images from data acquired along Cartesian, radial, and spiral trajectories.
• Achieved acceleration factors of up to 4, 6, and 4 for Cartesian, radial, and spiral trajectories, respectively.
• Demonstrated performance on par with state-of-the-art reconstruction techniques.
• Reconstruction error was found to be dependent on acceleration factor and sampling trajectory.

Conclusions:

• CRAUNN offers a versatile and effective solution for image reconstruction in rapid parallel MRI.
• The neural network approach provides robustness across diverse sampling patterns.
• The method holds potential for accelerating MRI scans without compromising image quality.