Highly undersampling dynamic cardiac MRI based on low-rank tensor coding.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Cardiovascular Research

Background:

• Dynamic cardiac magnetic resonance imaging (CMRI) is crucial for non-invasive cardiovascular disease assessment.
• Long acquisition times in CMRI limit its clinical utility due to high spatiotemporal resolution and whole-heart coverage requirements.
• Conventional methods treat CMRI data as 2D matrices, limiting exploitation of inherent data redundancy.

Purpose of the Study:

• To introduce a low-rank tensor coding (LRTC) model with tensor sparsity for compressive sensing (CS) in dynamic CMRI.
• To improve the efficiency and accuracy of dynamic CMRI reconstruction.
• To leverage redundancy in high-dimensional CMRI data for better image recovery.

Main Methods:

• Proposed a low-rank tensor coding (LRTC) model that treats groups of similar 3D patches as low-rank tensors.
• Utilized the alternating direction method of multipliers (ADMM) for model optimization, incorporating a soft threshold operator for L1 norm relaxation.
• Employed higher-order singular value decomposition (HOSVD) to process high-dimensional tensors and identify spatiotemporal correlations.

Main Results:

• The LRTC model effectively captures the sparse components of dynamic CMRI data.
• The method makes full use of redundancy between feature vectors of adjacent positions.
• Reconstruction accuracy was comparable to low-rank matrix recovery methods and superior to conventional sparse recovery methods on cardiac cine and myocardial perfusion datasets.

Conclusions:

• The proposed LRTC model offers an effective approach for compressive sensing in dynamic CMRI.
• This method enhances image reconstruction accuracy and efficiency, addressing limitations of conventional techniques.
• LRTC shows promise for accelerating dynamic CMRI acquisition while maintaining diagnostic image quality.