Synopsis

This lecture presents the main techniques to undersample MR spectroscopic imaging (MRSI) for increased imaging speed, including parallel imaging, compressed sensing and model-based subspace reconstruction

Objectives

• Describe the basics of the main techniques for undersampling, such as parallel imaging, compressed sensing and model-based reconstruction
• Study the application of undersampling techniques to MRSI for reduction of scan time, increase of spatial and spectral resolution, and volumetric coverage

Undersampled MRSI techniques

Conventional MRSI techniques acquire one data point for each voxel and spectral point to be reconstructed, which results in significantly long scan times. As a consequence, spatial and spectral resolution, as well as volumetric coverage are usually sacrificed for clinically feasible scan times. Undersampling represents a powerful means to overcome these limitations by acquiring fewer k-space points and exploiting some type of redundancy in the reconstruction process to avoid aliasing artifacts. Redundancies can be handmade, such as acquisition with multiple coils, or inherent to the data, such as compressibility. This lecture will review the basics of undersampling techniques and their application to MRSI. The different undersampling techniques will be grouped into three categories:

• Parallel imaging: different sensitivity profiles in a coil array are exploited to undersample k-space in a regular fashion. Standard SENSE and GRAPPA have been applied to phase-encoded MRSI (CSI) [1] and echo-planar MRSI [2]. Superresolution SENSE reconstruction that exploits intra-voxel coil sensitivities was also demonstrated for echo-planar MRSI [3]. The main limitation of parallel imaging is noise amplification.
• Compressed sensing: spatial and temporal correlations in the MRSI data can be exploited to impose sparsity in the reconstruction of randomly undersampled k-space data [4]. Compressed sensing and parallel imaging can be combined for higher acceleration rates [5]. Even though compressed sensing is very robust to noise, spatial and spectral blurring appears as a limiting factor.
• Model-based reconstruction: physical models of the MRSI signal defined by a few parameters can be exploited to reconstruct undersampled data. One of the most recent approaches in model-based MRSI known as SPICE uses the physical models to form sub-spaces that describe the whole MRSI in a compact fashion, and enable not only to accelerate, but to separate lipids from metabolites [6]. One of the limitations in model-based MRSI is the assumptions made to the model, which might produce leakage from one components onto the others.

Acknowledgements

References

[1] Dydak U et al. Magn Reson Med. 2001;46(4):713-22.

[2] Lin FH et al. Magn Reson Med. 2007; 57(2):249-57.

[3] Otazo R et al. Neuroimage. 2009; 47(1):220-30.

[4] Hu S et al. J Magn Reson. 2008;192(2):258-64

[5] Otazo R et al. ISMRM 2009; 331

[6] Lam F et al. Magn Reson Med. 2016;76(4):1059-70.