Benchmarking 3D Multi-Coil NC-PDNet MRI Reconstruction

Motivation: Deep Neural Networks (DNNs) show promise in reconstructing undersampled MRI data but face challenges with non-Cartesian 3D multi-coil acquisitions due to high memory demands and training complexity. Goal(s): To address the research gap in validating DNNs performances for non-Cartesian data in 3D multi-coil settings and assess undersampling patterns' impact on reconstruction quality. Approach: We extend the Non-Cartesian Primal-Dual Network (NC-PDNet) for multi-coil 3D reconstruction, evaluating different training configurations and benchmarking four non-Cartesian undersampling patterns. Results: NC-PDNet with GoLF-SPARKLING trajectory achieved gains of +2.43/+0.01 in PSNR/SSIM for 32-channel data, reconstructing a 1mm isotropic volume in 4.95 seconds using 5.49 GB GPU memory. Impact: Achieving fast, high-quality reconstructions with reasonable GPU memory usage demonstrates our approach's viability for clinical research. Benchmarking the GoLF-SPARKLING trajectory against established non-Cartesian baselines in 3D multi-coil settings validates its future application in more challenging experiments, notably at higher resolutions.