Dictionary learning and time sparsity for dynamic MR data reconstruction.

Area of Science:

• Medical Imaging
• Signal Processing
• Machine Learning

Background:

• Undersampled k-space data acquisition accelerates Magnetic Resonance Imaging (MRI).
• Compressed Sensing (CS) theory enables image reconstruction from undersampled data using sparsity models.
• Dictionary Learning (DL) offers improved performance over fixed basis transforms for sparse representations.

Purpose of the Study:

• To develop and evaluate an iterative algorithm for cardiac cine MRI reconstruction using Dictionary Learning (DL) with Cartesian undersampling.
• To enhance the reconstruction quality and convergence rate for dynamic MRI data.

Main Methods:

• An iterative algorithm applying DL to reconstruct cardiac cine MRI data from Cartesian undersampled k-space.
• Local processing of spatio-temporal 3D patches and independent treatment of real/imaginary parts.
• Incorporation of temporal gradients as an additional constraint to accelerate convergence.

Main Results:

• The proposed DL-based method systematically outperforms k-t FOCUSS, a successful CS method.
• Demonstrated improved reconstruction quality for dynamic MRI data, especially at high acceleration rates.
• The algorithm enables efficient application of DL for cardiac cine MRI.

Conclusions:

• Dictionary Learning provides a superior approach for reconstructing undersampled dynamic MRI data compared to fixed basis methods.
• The developed iterative algorithm effectively utilizes DL for cardiac cine MRI, offering significant improvements.
• Temporal gradient enforcement further enhances reconstruction performance and convergence speed.