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

Abstract

Purpose To reduce noise and artifacts in 23Na MRI with a Compressed Sensing reconstruction and a learned dictionary as sparsifying transform. Methods A three-dimensional dictionary-learning compressed sensing reconstruction algorithm (3D-DLCS) for the reconstruction of undersampled 3D radial 23Na data is presented. The dictionary used as the sparsifying transform is learned with a K-singular-value-decomposition (K-SVD) algorithm. The reconstruction parameters are optimized on simulated data, and the quality of the reconstructions is assessed with peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The performance of the algorithm is evaluated in phantom and in vivo 23Na MRI data of seven volunteers and compared with nonuniform fast Fourier transform (NUFFT) and other Compressed Sensing reconstructions. Results The reconstructions of simulated data have maximal PSNR and SSIM for an undersampling factor (USF) of 10 with numbers of averages equal to the USF. For 10-fold undersampling, the PSNR is increased by 5.1 dB compared with the NUFFT reconstruction, and the SSIM by 24%. These results are confirmed by phantom and in vivo 23Na measurements in the volunteers that show markedly reduced noise and undersampling artifacts in the case of 3D-DLCS reconstructions. Conclusion The 3D-DLCS algorithm enables precise reconstruction of undersampled 23Na MRI data with markedly reduced noise and artifact levels compared with NUFFT reconstruction. Small structures are well preserved.