On retrospective k-space subsampling schemes for deep MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Accelerated MRI acquisitions reduce scan time but often yield imprecise reconstructions, particularly with conventional Cartesian-rectilinear subsampling.
• Deep Learning (DL) models show promise for improving accelerated MRI reconstruction quality, but performance can be limited by the subsampling scheme.
• Non-rectilinear and non-Cartesian trajectories offer alternative k-space subsampling strategies in MRI.

Purpose of the Study:

• To investigate the impact of various k-space subsampling schemes on the quality of accelerated MRI reconstructions using Deep Learning (DL) models.
• To compare the performance of DL-based reconstructions trained and evaluated on different subsampling strategies, including Cartesian-rectilinear, Cartesian non-rectilinear, and non-Cartesian trajectories.
• To evaluate reconstruction quality across scheme-specific and multi-scheme training frameworks.

Main Methods:

• Utilized the Recurrent Variational Network (RecurrentVarNet) as the DL-based MRI reconstruction architecture.
• Retrospectively subsampled fully-sampled multi-coil k-space data from three datasets using eight distinct schemes (four Cartesian-rectilinear, two Cartesian non-rectilinear, two non-Cartesian) at various acceleration factors.
• Conducted experiments in both scheme-specific (model trained/evaluated per scheme) and multi-scheme (single model trained on all schemes) frameworks.

Main Results:

• RecurrentVarNets trained and evaluated on non-rectilinearly subsampled data demonstrated superior reconstruction performance compared to rectilinearly subsampled data, especially at high acceleration factors.
• In the multi-scheme setting, reconstruction performance for rectilinearly subsampled data improved when compared to scheme-specific experiments.
• Non-rectilinear subsampling schemes consistently yielded higher quality reconstructions across both experimental frameworks.

Conclusions:

• Non-rectilinear k-space subsampling schemes, when used with DL-based reconstruction methods like RecurrentVarNet, significantly enhance accelerated MRI image quality.
• DL models trained on non-rectilinearly subsampled data offer a promising approach to optimize MRI scan time while maintaining or improving image quality.
• The findings highlight the importance of selecting appropriate k-space trajectories for effective DL-based accelerated MRI reconstruction.