Sparse Quantification of 1H-MRS Based on Metabolites Profiles in Time- Frequency Employing Pursuit Algorithm; A Phantom Study

Background:

Performance of quantification methods utilized for proton magnetic resonance spectroscopy (1H-MRS), implemented either in time-domain or in frequency-domain, is limited due to static field (B0) inhomogeneities and the overlapping nature of metabolites in actual low-SNR environments. Sparse representation methods for MRS quantification have suggested robust and high performance algorithms which have been previously implemented based on Gaussian and Lorentzian models, to be selected by some pursuit methods, e.g. parallel basis selection method based on the focal underdetermined system solver (FOCUSS) algorithm. FOCUSS algorithm performs well in correlated environments, however it is computationally more intense compared to the selection methods.

Objectives:

Here, we proposed a sparse quantification method for 1H-MRS in time-frequency domain, achieved by continues wavelet transform (CWT), and employing sparse features of the whole simulated metabolites spectra existing in frequency, as well as properties of profiles in time. Stability and accuracy of the proposed technique was confirmed by simulated and phantom data, resulting in correct quantification of the metabolites of interest in 1H-MRS signals of brain.

Patients and Methods:

Phantom preparation- An MRS phantom with 13 vials, filled with solutions containing the most important brain metabolites with known concentration was used to simulate tumorous and normal brain tissues; 5 vials were filled with pure metabolites in order to achieve metabolites profile, and the other vials contained relevant concentrations of Alanine, Choline Chloride, Creatine, Myo-Inositol and N-Acetyl-L-Aspartic Acid, mimicking the tumorous and normal human brain tissues. Data acquisition- Proton MRS imaging experiments were performed on a 1.5T (Siemens Avanto) MRI/MRS system in the room temperature using Point REsolved Spectroscopy (PRESS) pulse sequence with manufactures built-in auto-shimming on the volume-of-interest, CHESS water suppression and 3D imaging parameters as follows: TE/TR = 30/1500 ms, voxel size = 101010 mm3, frequency bandwidth = 1200 Hz and number of data- points = 1024. Quantification- Eq. 1 expresses MRS signal as a function of linear combination of metabolites profile with added baseline and noise, in which K is the number of metabolites in the signal, m (k,n) represents k-th metabolite information profile, and a(k) is amplitude (weight) of each profile in the signal. The term B (n) is the baseline signal generated by macromolecules and e (n) is white Gaussian noise. Block diagram of the quantification procedure is shown in, with details as follows: 1) Employed metabolite dictionary can be created either from simulated profile or from phantom (to be adjusted for frequency shift, time shift, receive coil inhomogeneity and unwanted signals); 2) Complex Morlet CWT (Eq. 2) was applied to both signal and metabolites profiles (simulated-based or phantom-based); 3) Noise estimation was performed based on a rough approximation of each metabolite, being applied onto the dictionary and the signal; and 4) Sparse quantification employing FOCUSS Pursuit algorithm estimated the sparse representation of the signal with respect to the constructed dictionary.

Results:

PRMSE of the quantification for 5 metabolites was shown in three types of signals: 1) Simulated signal with infinite signal-to-noise ratio (SNR), quantified by the simulated dictionary; 2) Simulated signal with SNR = 10 dB, quantified by the simulated dictionary; and 3) Signal acquired from 7 different types of solutions in the phantom, quantified by the phantom-based dictionary.

Conclusions:

Results show that the proposed method can quantify metabolites of 1H-MRS signal with low level of error (< 1% for simulated signal (SNR = infinite), < 3% for simulated signal (SNR = 10) and < 9% for phantom signal with the dictionary based on phantom). The proposed procedure finds the sparse representation of the signal after exploiting almost all information of the signal in the linear sparse combination of a number of dictionary profiles after being transformed onto the time-frequency domain. Using metabolites profile acquired from the phantom also added to the accuracy of metabolites estimation.