The recently suggested approach of magnetic resonance fingering (MRF) relies on matching the response of a system undergoing semi-random excitations to a dictionary of pre-simulated signal curves with fixed parameters¹ (T₁, T₂, proton density PD, transmit inhomogeneity B₁₊). The closest matching dictionary entry is used to estimate the system’s parameters.

The application of MRF to magnetic resonance spectroscopy (MRS) requires careful planning. A full MRS spectrum contains over a dozen metabolites; even with only a single spin-1/2 system with three unknown parameters (T₁, T₂¸ B₁₊), a typical dictionary contains hundreds of thousands of entries. Furthermore, most metabolites are described by complex Hamiltonians which are significantly more time consuming to simulate than a single spin-1/2 system. The macromolecular baseline is also difficult to model and is often fitted non-parametrically. Acquisitions cannot be made arbitrarily short, as at least ~ 0.5-1.0 seconds are required for water suppression and the sampling of the free induction decay. Finally, the very low signal to noise ratio (SNR) of MRS makes temporal fitting extremely unreliable.

A fingerprinting scheme to accurately quantify the T1, PD and B1+ of all metabolites’ is presented. To overcome the low SNR, we averaged subsets of measurements together; this allowed for fitting each averaged subset to pre-simulated basis sets using conventional fitting algorithms. Varying the repetition time TR and the excitation angle a introduced T1 and B1+ weighting. Standard match and pursuit algorithms were used to fit averaged signal curves to average dictionary entries.

Data acquisition: A single 1.5×1.5×2 cm³ voxel was placed in the white matter of a healthy volunteer (Fig. 1a) at 3T (Tim Trio, Siemens) and a PRESS sequence was acquired with 120 averages. TE was fixed at 35 ms. TR and α, the flip angle of the first PRESS pulse, were varied per-scan as shown in Fig. 1b, 1c. A moving N_avg=8 scan average was applied to the data, i.e. the i^th scan was replaced by the sum of scans {i, i+1, ..., i-1+N_avg} to increase SNR. The N-acetyl-aspartate (NAA), choline (Cho) and creatine (Cr) peaks were fitted with VESPA⁶. A correction (based on simulations of the Bloch equations) was applied to each scan's intensity, since varying the flip angle α distorts the excitation pulse's profile slightly. A non-suppressed water acquisition was acquired for quantification.

Dictionary generation and entry matching: Only singlets were simulated for this preliminary study. Echo intensities for each scan were computed by solving the Bloch equations, and were summed with the same moving average. Prior to matching all dictionary entries and experimental data had their initial point normalized to unity. Dictionary matching was achieved with a standard match & pursuit algorithm².

Absolute Quantification: Following matching and determination of T₁, B₁₊, the dictionary entry was rescaled to fit the experimental time series to extract its PD. The relative intensity of each dictionary curve, compared to a standard full excitation with TR=2000 ms, α=90°, was also calculated from simulation and used to perform absolute quantification using the reference water scan using literature values⁷ of T₂.

"Gold Standard": PD was estimated using a PRESS with TR/TE=2000/35, 90 averages. T₁ was estimated using standard inversion recovery with 5 inversion points. B₁₊ was estimated using a saturated double angle method³. To ensure fair comparison, total time of all sequences was kept equal to the stochastic excitation sequence.

Fig 2a shows a sample 8-scan average spectrum. Fig. 2b shows the averaged NAA level as a function of scan number after averaging (blue) and the closest matching dictionary entry (red). Fitted concentrations, T1 and B1+ values are summarized in Fig. 3. White matter concentrations and T₁ values agree with the literature^4,5. The conventional and MRF approaches differ by about 10%-20%. It is difficult to ascertain whether one method is superior; phantom experiments (not shown) have exhibited smaller discrepancy levels ~5%. A good consistency check is afforded by B₁₊, which shows excellent agreement between all metabolites using MRF.

The principles off fingerprinting can be successfully adapted to single voxel spectroscopy if appropriate modifications to the basic scheme are made:scans must be averaged, fit and only then matched to corresponding dictionary entries.The number of scans in the moving average should be optimized: on the one hand, increasing N_avg increases the SNR and reduces fitting error; on the other hand, it diminishes the signal variability introduced by the stochastic excitation and, consequently, its dependence on T₁ and B₁₊. Note that T₂ was not estimated since the long TRs used in MRS greatly diminish transverse coherences and the signal's dependence on T_2.