Magnetic resonance spectroscopy (MRS) is a commonly used technique for clinical studies investigating various (mainly neurological) disorders. Its main limitations are SNR, spectral overlap and sensitivity to various artefacts. SNR limits achievable resolution and requires long acquisition times. Spectral overlap means that from the ~20 potentially detectable metabolites, people often focus on the three predominant singlets (NAA, creatine and choline). Sensitivity to artefacts means that great care has to be taken during acquisition to ensure proper water and fat suppression, good shimming, and lots of other things. Mainly two approaches exist to get more out of spectroscopy: increase spatial information by going to chemical-shift imaging (CSI); increase spectral information by going to two-dimensional spectroscopy (e.g. 2D JPRESS or COSY), editing (e.g. difference editing or multiple quantum coherence filtering), higher field strengths (e.g. 7T) and/or better shimming [1]. This educational talk will focus on 2D spectroscopy, more specifically on how to extract the maximum amount of information from such data. In this text, the wording 1D and 2D will always denote spectral dimensions, and NOT spatial dimensions as used in CSI.

Many of the ~20 metabolites in proton brain MRS resonate at similar chemical-shift (CS) frequencies [1]. Additionally, most metabolites do not have a single singlet peak, but are split up into multiplets by J-coupling. Hence, a typical 1D spectrum is overcrowded with many peaks overlapping each other. It is difficult to separate for instance glutamate and glutamine at 3T. In order to detect metabolites and separate their signal contributions from each other, it is normally required to modify the acquisition to either filter out unwanted resonances or to encode more spectral information by 2D spectroscopy. The advantage of filtering is that the acquisition can be highly optimised for a single metabolite, while the advantage of 2D MRS is that the whole range of metabolites can potentially be detected in a single scan. The most common 2D techniques applied to in vivo MRS are 2D J-resolved spectroscopy (JPRESS) and 2D COSY [2]. JPRESS uses the PRESS sequence (90°-180°-180° with slice selection in three orthogonal directions), while COSY uses a (partly) STEAM based acquisition incorporating at least two 90° pulses (out of the three pulses required for localising a voxel). JPRESS therefore exhibits higher SNR, while J-coupled peaks are potentially better separated by COSY. Main advantage of JPRESS is probably its robustness, especially when sampling the maximum echo [3].

The most accurate quantification methods for determining individual metabolite concentrations generally imply spectral fitting. The physical properties (number of spins, CS frequencies, J-couplings) of in vivo brain metabolites are well characterised and the resulting spectra can be either simulated or measured directly in pure phantom solutions. These individual metabolite spectra can be used as basis set in the fitting function for determining the individual metabolite concentrations. Fitting a measured spectrum to the full individual metabolite basis sets incorporates the full amount of prior information in the quantification, hence yielding potentially the most accurate concentration values. For 1D MRS quantification, the quasi golden standard is the commercially available LCModel program (LCMODEL Inc., Oakville, ON, Canada) [4]. A non-linear fit determines line-broadening, small shifts, baseline, phases and line-shape distortions, while the concentrations are determined using these parameters in a linear least squares fit doing a “Linear Combination of Model Spectra”. Absolute quantification yielding concentrations in mM (=1 mol/m³) requires references (such as internal water, external reference phantom, RF), is challenging and prone to errors. Therefore, concentrations are typically given as ratios to creatine or water.

Different spectral fitting methods exist for 2D MRS, with the simplest ones being peak fitting inherited from the NMR field. Another way is to look at a single cross-section through the 2D spectrum, hence generating a pseudo 1D spectrum (e.g. called TE-averaged MRS), and applying common 1D fitting methods such as LCModel to it (Fig. 1). Alas, a lot of spectral information is not considered by doing this, and the only way to extract the maximum information out of the 2D spectra is by a linear combination of the full 2D model spectra. There is no commercially available program existing, but various methods are described in the literature. The most widely used program is called Prior-knowledge Fitting (“ProFit”) [5], which is available for download [6]. It was originally developed to quantify JPRESS brain spectra, but in the meantime it got adapted to other organs (prostate) [7] as well as other acquisition methods (COSY) [8]. It is not as automated and easy to use as LCModel.

ProFit consist of a non-linear least squares fit to determine lineshape (linewidth + distortions), (small) shifts and phases. These are used to update data or basis functions, in order to determine concentrations in a linear least-squares fit using the 2D metabolite spectra as basis function for a linear superposition. In order to decrease computational complexity, the fits are repeated three times with increasing complexity (first only singlets, final step all metabolites), initialising the next iteration with prior results (Fig. 2). Baseline distortions were neglected in the initial, first ProFit version, but were included into a second version (along with improved line modelling), thus reducing contamination from macromolecules and hence improving quantification, at the cost of increased computational complexity [9]. Basis spectra were simulated with published metabolite CS-shifts and couplings [10] using GAMMA [11], a quantum-simulation library developed for NMR. It is possible to detect and quantify more brain metabolites as compared to standard 1D PRESS.