Deep low-Rank plus sparse network for dynamic MR imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Dynamic magnetic resonance (MR) imaging often relies on low-rank plus sparse (L+S) decomposition, but parameter selection is empirical and acceleration is limited.
• Existing deep learning methods for compressed sensing MR imaging (CS-MRI) reconstruction rarely incorporate low-rank priors.

Purpose of the Study:

• To propose a model-based low-rank plus sparse network (L+S-Net) for dynamic MR reconstruction.
• To address limitations in parameter selection and acceleration rates of current CS-MRI methods.

Main Methods:

• Developed L+S-Net using an alternating linearized minimization method for optimization with low-rank and sparse regularization.
• Incorporated learned soft singular value thresholding for component separation.
• Unrolled iterative steps into a network with learnable regularization parameters.

Main Results:

• Demonstrated global convergence of L+S-Net under standard assumptions.
• Achieved superior performance compared to state-of-the-art CS and deep learning methods on cardiac cine datasets.
• Showcased potential for extremely high acceleration factors (up to 24×).

Conclusions:

• L+S-Net offers a robust and efficient solution for dynamic MR reconstruction.
• The model enables significant acceleration in MR imaging without compromising image quality.
• This approach holds promise for advancing dynamic MR imaging applications.