MR spectroscopy is traditionally performed in the clinical environment using a single-voxel (SV) approach (1). On occasion, 2D or 3D MR spectroscopic imaging (MRSI) may be performed in conjunction with (usually) PRESS-based localization (2,3). The PRESS sequence is usually performed to restrict the signal localization to the sub-region of the brain of interest, which allows smaller fields-of-view to be used, avoids artifacts from peripheral regions, and reduces the number of spectra that need to be processed. However, this approach also suffers several drawbacks, including lack of coverage of potentially important brain regions, chemical shift displacement artifacts at the edge of the PRESS voxel (4), and dependence on the operator of correct voxel placement. For these reasons, over the years there has been development of alternative MRSI techniques which either have whole-slice (5), multi-slice (6), or 3D whole-brain coverage (7). Whole-brain coverage at high spatial resolution should be the goal for clinical applications (since, a priori, the location and extent of a metabolic abnormality may not be known ahead of time), but important challenges to this methodology are the need to perform it in clinically reasonable scan times, to optimize and simplify data analysis, quantitation, and visualization software. In addition, recognizing and removing artifacts is an important step for clinical implementation in order to avoid incorrect interpretation by clinicians. Because extended spatial coverage at high resolution leads to long encoding-times, high speed acquisition techniques such as echo-planar spectroscopic imaging (EPSI) are required (5), preferably also in combination with parallel imaging acceleration such as SENSE or GRAPPA (8-10). Also, high-resolution multi-dimensional data sets (particularly those recorded with 32-channel coils) are large (e.g. > 5GB not unusual) so attention equally needs to be paid to acceleration of the reconstruction and data transfer.

In general, MRSI is more sensitive to imperfections and artifacts than conventional MRI, because the signals to be measured are small, often contaminated by much larger unwanted artifact signals (water and lipid), and also has a much more stringent requirement on field homogeneity. Field inhomogeneity may arise from several reasons, including the air-tissue interfaces within the cranium, post-surgical effects, hemorrhage, or incorrect adjustment of shim currents. Generally, this will manifest as decreased signal on MRSI, but may occasionally give increased signal if it leads to residual water or lipid contamination in the spectrum. Importantly, various ‘quality’ measures may be derived from the MRSI data itself which can help in interpreting metabolic images and deciding if focal signal intensities are real or artifact (11). Such quality measures may include metabolite or water peak linewidth, water peak intensity, or an uncertainty measure from spectral curve-fitting (e.g. Cramer-Rao Lower Bounds (CRLB)). Finally, atlas-based analysis methods and integration of spectra across voxels may improve spectral quality (particularly signal-to-noise ratio (SNR)) compared to analysis of individual voxels from the dataset.