Dynamic MRI of the abdomen using parallel non-Cartesian convolutional recurrent neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Undersampled non-Cartesian abdominal dynamic parallel MRI data presents challenges in image quality and reconstruction time.
• Existing methods like XD-GRASP, L+S, BCS, and 3D CNN have limitations in handling complex motion and achieving high fidelity reconstructions.

Purpose of the Study:

• To develop a deep learning approach for improving image quality and reducing computational time in reconstructing undersampled non-Cartesian abdominal dynamic parallel MR data.
• To leverage spatial and temporal redundancy for enhanced data fidelity in MR image reconstruction.

Main Methods:

• Development of parallel non-Cartesian convolutional recurrent neural networks (PNCRNNs).
• Evaluation of PNCRNNs against state-of-the-art algorithms (XD-GRASP, L+S, BCS, 3D CNN) using various acceleration rates and motion patterns.
• Assessment of reconstruction performance across different imaging applications.

Main Results:

• PNCRNNs achieved significant peak SNR improvements: 9.07 dB over XD-GRASP, 9.26 dB over L+S, 3.48 dB over BCS, and 3.14 dB over 3D CNN at R=16.
• Reconstruction time was reduced to 18 ms per bin, two orders of magnitude faster than XD-GRASP, L+S, and BCS.
• Excellent reconstruction quality was demonstrated for diverse motion patterns, k-space trajectories, and imaging applications.

Conclusions:

• The proposed PNCRNN algorithm offers substantial improvements in image quality for dynamic abdominal imaging compared to existing methods.
• PNCRNNs achieve rapid reconstruction speeds (up to 50 bins/sec) due to an efficient Toeplitz approach.
• This deep learning method represents a significant advancement for dynamic parallel MRI reconstruction.