Learned Low-Rank Priors in Dynamic MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning methods show promise in dynamic MR cine imaging but often overlook the low-rank prior.
• Existing methods primarily rely on sparsity priors, potentially limiting reconstruction performance.

Purpose of the Study:

• To introduce a novel deep learning network, SLR-Net, that leverages both sparse and low-rank priors for dynamic MR imaging.
• To improve the reconstruction quality and robustness of dynamic MR cine imaging.

Main Methods:

• Proposed a learned singular value thresholding (Learned-SVT) operator to exploit low-rank priors.
• Developed a model-based unrolling sparse and low-rank network (SLR-Net) based on the iterative shrinkage-thresholding algorithm (ISTA).
• Evaluated SLR-Net in single-coil and multi-coil scenarios, including real-time datasets.

Main Results:

• SLR-Net outperformed state-of-the-art compressed sensing and sparsity-driven deep learning methods in single-coil reconstructions.
• Demonstrated strong robustness to various undersampling patterns.
• Achieved excellent results in multi-coil scenarios under high acceleration and showed capability in real-time applications.

Conclusions:

• SLR-Net effectively integrates sparse and low-rank priors for superior dynamic MR image reconstruction.
• The proposed method offers enhanced performance, robustness, and flexibility for dynamic and real-time MRI applications.