Wavelet-based Techniques in MRS

Amagnetic resonance spectroscopic (MRS) signal is made of several frequencies typical of the active nuclei and their chemical environments. The amplitude of these contributions in the time domain depends on the amount of those nuclei, which is then related to the concentration of the substance (Hornak, 1997). This property is exploited in many applications of MRS, in particular in the clinical one. The MRS spectra contain awealth of biochemical information characterizing themolecular content of living tissues (Govindaraju et al., 2000). Therefore, MRS is a unique non-invasive tool for monitoring human brain tumours, etc. (Devos et al., 2004), if it is well quantified. When anMRS proton signal is acquired at short echo-time (TE), the distortion of spectral multiplets due to J-evolution can be minimized and the signals are minimally affected by transverse relaxation. Such signals exhibit many more metabolite contributions, such as glutamate and myo-inositol, compared to long TE spectra. Therefore, an MRS signal acquired at short TE presents rich in vivometabolic information through complicated, overlapping spectral signatures. However, it is usually contaminated by water residue and a baseline which mainly originates from large molecules, known as macromolecules. As the shape and intensity of the baseline are not known a priori, this contribution becomes one of the major obstructions to accurately quantify the overlapping signals from the metabolites, especially by peak integration, which is commonly used in frequency-based quantification techniques. Also, by seeing only the frequency aspect, one loses all information about time localization. A number of quantification techniques have been proposed, which work either in the time domain (see Vanhamme et al. (2001) for a review) or in the frequency domain (see Mierisova & Ala-Korpela (2001) for a review). The time-domain based methods are divided into two main classes: on one side, non-interactive methods such as SVD-based methods (Pijnappel et al., 1992) and, on the other side, methods based on iterative model function fitting using strong prior knowledge such as QUEST (Ratiney et al., 2004; 2005), LCModel (Provencher, 1993), AQSES (Poullet et al., 2007), or AMARES (Vanhamme et al., 1997).