NC-PDNet: A Density-Compensated Unrolled Network for 2D and 3D Non-Cartesian MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning shows promise for magnetic resonance image (MRI) reconstruction.
• Existing methods often struggle with non-Cartesian acquisition settings.

Purpose of the Study:

• To develop and validate an unrolled neural network for density-compensated (DCp) non-Cartesian MRI reconstruction.
• To assess the performance and generalizability of the proposed network against baseline models.

Main Methods:

• Design of NC-PDNet (Non-Cartesian Primal Dual Network), the first DCp unrolled neural network.
• Ablation study to validate key components of NC-PDNet.
• Generalizability experiments using out-of-distribution data (e.g., knee to brain).
• Comparison with baseline models like U-Net and Deep Image Prior.

Main Results:

• NC-PDNet outperforms baseline models visually and quantitatively across all tested settings.
• Achieved up to 1.2 dB improvement in peak signal-to-noise ratio (PSNR) for 2D multi-coil acquisitions.
• Demonstrated at least 1 dB PSNR gain in generalization settings.
• Open-source implementation provided, including Non-uniform Fourier Transform in TensorFlow.

Conclusions:

• NC-PDNet is an effective deep learning approach for non-Cartesian MRI reconstruction.
• The network shows robust performance and excellent generalizability.
• The proposed model advances the field of accelerated MRI acquisition and reconstruction.