A parallel MR imaging method using multilayer perceptron.

Area of Science:

• Magnetic Resonance Imaging
• Machine Learning
• Image Reconstruction

Background:

• Undersampled k-space data in MRI leads to aliasing artifacts, necessitating robust reconstruction techniques.
• Accelerating MRI acquisition is crucial for reducing scan times and improving patient comfort.
• Traditional reconstruction methods may struggle with complex undersampling patterns.

Purpose of the Study:

• To develop a fast and accurate MR image reconstruction method using multilayer perceptron (MLP).
• To address aliasing artifacts arising from k-space subsampling.
• To enable real-time image reconstruction and accelerate MRI acquisition.

Main Methods:

• Applied MLP algorithm trained on data to map aliased k-space images to alias-free images.
• Utilized multichannel real and imaginary images from subsampled k-space as MLP input.
• Processed aliased data line-by-line using the learned MLP architecture for artifact reduction.

Main Results:

• Achieved superior image reconstruction quality compared to existing methods, indicated by lower normalized root-mean-square error.
• Demonstrated applicability to various k-space subsampling patterns in the phase encoding direction.
• Facilitated parallel processing for further reduction in reconstruction time.

Conclusions:

• Proposed a novel machine learning-based reconstruction method for accelerated MRI.
• Successfully reconstructed high-quality MR images from subsampled k-space data.
• Highlighted the method's flexibility with sampling patterns and real-time reconstruction capabilities.