Step 1: Spectral fitting to a specific model

The data quantification procedure discussed below assumes that the data was collected with the known experimental parameters and sequence components. The data should be pre-processed prior to quantification step (e.g. multiple channel reconstruction, eddy current correction, scan to scan frequency/phase correction due to motion, frequency drift correction from gradient heating, etc). Therefore, the data is assumed to be as close to the theoretical model as possible. The first step in the fitting procedure is deciding whether time- or frequency domain will be chosen. Although two domains are theoretically equivalent (because FFT is a linear operation), the choice of the quantitation method depends on the data under investigation and the pursued objectives. One of the most popular free software routines that offers both time and frequency domain fitting options is JMRUI (1). The spectral fitting algorithms (both in the time- and frequency domains) can be roughly divided into two main categories: black-box and methods based on iterative model function fitting, which allow the incorporation of constraints and prior knowledge (for a comprehensive review of time-and frequency-domain fitting, see Poullet et al (2)). While black-box methods can provide reasonable estimations in some specific cases, the iterative methods with the use of the prior knowledge are required in most cases to obtain biologically useful information (3). Thus iterative approaches that use parametric models are generally favored over black box methods. The spectrum can be modeled as the Fourier transform of the sum of decaying sinusoids, for each metabolic compound. The fixed values for spectral intensity and phases for a particular frequency location (typically in units of PPM) are obtained from an a priori spectral information, and they are collectively referred to as “prior knowledge”. A popular commercial package LCModel utilizes such prior knowledge in the form of linear combination (LC) of isolated compound spectra (4,5). These metabolic spectra are referred to as “basis sets”. Well defined metabolite prior information results in more consistent and complete estimations of the actual data. And visa versa - the incorporation of incorrect prior knowledge will typically lead to systematic bias. As a consequence, a particular care is needed to obtain such a basis set, which is typically done via phantom solution measurements or through numerical simulations. An advantage of the phantom measurement method is that additional experimental imperfections are reflected in the models (e.g. spatial profiles, sequence imperfections, etc). However there are many more disadvantages for this method – expensive chemicals, time consuming preparation of 15-25 pH balanced metabolites, specific sequence/parameter dependence, achieving uniform physiological temperature for many hours/days of the data acquisition. Alternatively, the creation of prior information for spectral analysis and fitting routines can be done via the density matrix based numerical simulations. The major consideration in the case of basis set generations is over how much prior information needs to be included to achieve usable metabolite models via numerical simulations. Often the expected SNR levels and typical data quality of clinical MRS data preclude the need for the use of more complicated simulations. Many of the spectral analysis models for clinical PRESS, STEAM and spin-echo MRS metabolite data from lower field MRI scanners (e.g. 1.5 Tesla) are created using simulations with ideal RF pulse objects. However, at higher magnetic fields (≥ 3 T), additional levels of physics or programming complexity are necessary to accurately simulate the MRS data. This is particularly true for high SNR single voxel data and metabolite edited MRS data with longer echo time (TE > 20 ms). With increasing magnetic fields, the limited bandwidths of the volume selective RF pulses create spatial distortions which alter the resultant amplitudes and phases. For such cases, full density matrix calculations with 3D localization using experimental pulses create more accurate basis sets (6). Other important aspects of obtaining better spectral fitting results using simulation models include the completeness of the prior knowledge and more precise values of J-coupling constants/chemical shifts. The first point is an important requirement to have a full set of model spectra for fitting, otherwise the excluded models will affect the amplitude of the metabolites of interest. The second aspect deals with incorrect spectra generation due to wrong J-values/chemical shifts used for the simulations, especially for longer TE simulations and higher magnetic fields.

Step 2: Conversion of fitting results to absolute concentration

An internal or external reference is required to obtain the absolute concentration. Usage of the internal reference (e.g. total creatine or water) is arguably the most straightforward pathway towards the absolute quantification, but it requires an assumption that the reference concentration is known. External reference examples include phantom measurements and detection of electronic signal synthesized by an external device. Additional calibration factors should be calculated depending on the type of reference chosen. These factors include, but not limited to: relaxation differences (both T1 and T2), B1 and Bo corrections, NOE effects, localization, voxel composition and spectral visibility. Some of these factors are difficult to estimate and often require additional data acquisition. As a result, quantification issues should be carefully considered prior to data acquisition, especially for the large study. In certain cases, tissue segmentation using additional high resolution images and B1/Bo maps may be required. While the use of the brain water as an internal reference appears to be the most robust and convenient solution, there is still no gold standard during this conversion step, and as such the term “absolute quantification” should be used with caution.

Acknowledgements

No acknowledgement found.