Sparse and dense hybrid representation via subspace modeling for dynamic MRI.

Area of Science:

• Medical Imaging
• Signal Processing
• Computer Vision

Background:

• Compressed sensing and low-rank matrix recovery are key in dynamic magnetic resonance imaging (dMRI).
• Current methods often use separate formulations for sparse and low-rank priors.
• A unified approach is needed for joint sparse and low-rank modeling in dMRI.

Purpose of the Study:

• To develop a novel sparse and dense hybrid representation (SDR) model for dMRI.
• To unify the modeling of sparse and low-rank properties in a single framework.
• To improve reconstruction and separation performance in dMRI.

Main Methods:

• A unified sparse and dense hybrid representation (SDR) model is proposed.
• The model utilizes a learned dictionary of temporal basis functions.
• Spatial coefficients are modeled using Laplacian and Gaussian priors in two subspaces.
• An alternating direction algorithm is employed for efficient model solving.

Main Results:

• The SDR model effectively combines sparse and low-rank properties.
• Experiments demonstrate superior reconstruction and separation compared to state-of-the-art methods.
• The proposed algorithm efficiently solves the SDR model.

Conclusions:

• The novel SDR model offers a unified approach for joint sparse and low-rank dMRI.
• The developed algorithm is efficient and effective for the proposed model.
• This work shows significant potential for advancing dMRI reconstruction and separation techniques.