Three-dimensional dictionary-learning reconstruction of (23)Na MRI data

Area of Science:

• Medical Imaging
• Biophysics
• Signal Processing

Background:

• Sodium-23 (23)Na MRI is crucial for in vivo tissue characterization.
• Undersampling in (23)Na MRI significantly reduces scan time but introduces noise and artifacts.
• Compressed Sensing (CS) reconstruction techniques offer potential for high-quality imaging from undersampled data.

Purpose of the Study:

• To develop and evaluate a novel Compressed Sensing (CS) reconstruction algorithm for (23)Na MRI.
• To reduce noise and artifacts in (23)Na MRI by employing a learned dictionary as a sparsifying transform.
• To enhance the precision and quality of (23)Na MRI reconstructions from undersampled data.

Main Methods:

• A three-dimensional dictionary-learning compressed sensing (3D-DLCS) algorithm was developed for reconstructing undersampled 3D radial (23)Na data.
• A K-singular-value-decomposition (K-SVD) algorithm was used to learn the dictionary for the sparsifying transform.
• Reconstruction quality was assessed using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) on simulated, phantom, and in vivo (23)Na MRI data.

Main Results:

• The 3D-DLCS algorithm achieved maximal PSNR and SSIM at a 10-fold undersampling factor.
• Compared to Nonuniform Fast Fourier Transform (NUFFT) reconstruction, 3D-DLCS improved PSNR by 5.1 dB and SSIM by 24% for 10-fold undersampled data.
• Phantom and in vivo results confirmed markedly reduced noise and undersampling artifacts with 3D-DLCS, preserving small structures.

Conclusions:

• The 3D-DLCS algorithm provides precise reconstruction of undersampled (23)Na MRI data.
• This method significantly reduces noise and artifacts compared to conventional NUFFT reconstruction.
• The 3D-DLCS technique effectively preserves small anatomical structures in (23)Na MRI.