Comparison of Methods for Baseline Characterization of in vivo ‘H MR Spectra B. J. Soher, K. Young and A. A. Maudsley DVA Medical Center, MR Unit (114M), 4150 Clement St., San Francisco, CA 94121. INTRODUCTION Previous reports have shown that analysis of in vivo spectra benefits from the use of a priori metabolite information in a parameterized model (l-2). However, these spectra often contain broad uncharacterized signal contributions from residual water, lipids, and macromolecules, and accurate analysis requires that these baseline and metabolite signals be differentiated. Due to the complexity of creating an accurate model, baselines are often simply parameterized, based on assumptions of smoothness. In this report, we compare four methods for accounting for baseline contributions using either splines or wavelets. An iterative method, combining a non-parametric characterization of baseline signals with a parametric model of metabolites, was compared against a fully parameterized metabolite and baseline spectral analysis algorithm for both wavelet and spline methods respectively. The incorporation of a priori knowledge of baseline signals for each method is discussed. METHODS Metabolites were fitted parametrically in all methods (2). The iterative methods alternated optimizing the metabolite model and non-parametrically fitting the baseline over 4 or 5 iterations. The parametric methods created a parametric representation of the baseline using wavelet or spline coefficients and optimized it together with the metabolite model. The L-BFGS-B routine was used to optimize all models in both the iterative and parameterized methods. The iterative wavelet method is described in more detail elsewhere (2). For this study, a shift invariant, wavelet (SIW) filter was developed to improve the fit of data sections misaligned with the wavelet dyad of a scale most appropriate to its fit. The SIW filter performs 2N+l wavelet baseline characterizations for +N point shifts of the data. The final result is formed by taking the mean of the 2N+l values at each data location. N was set to one half of the minimum dyad size allowed by the metabolite linewidth times the wavelet scale multiplier. The scale multiplier was typically set to 4 times the metabolite linewidth to maintain a broadness separation between metabolite and baseline signals. All spline methods used fixed knot B-splines. Knot spacing was set at twice the estimated metabolite linewidth to maintain a broadness criterion. The iterative B-spline method used an external call to the IMSL Fortran routine BSLSQ to fit a set of fixed knot B-splines to the data. In the parametric methods, the baseline was modeled by either wavelet or spline coefficients, with a fixed set of dyads or spline knot placements. Initial wavelet coefficients were determined by setting the scale multiplier to 4 and doing a non-parametric fit to the data. Similarly, initial B-spline coefficients were obtained using the BSLSQ routine. A Monte Carlo analysis was performed for each of the four methods, using 100 repetitions. Peaks approximating singlets from NAA, Cr and Cho, were overlaid on a smoothed baseline from a short TE data acquisition. Noise was added to achieve a 1O:l SNR for the metabolites. Fig. la shows a sample spectrum, which has been analyzed using standard wavelet analysis (lb) and the improved SIW filter (1~). Both plots include the original and fitted baselines, and their difference. Chi square values were calculated for the difference between the fitted and original baselines. Baseline fits were further divided into metabolite and lipid regions to the left and right of 1.6 ppm respectively. Table 1 contains the averaged chi square results for the overall, metabolite, and lipid regions for all four methods. 6) 4.0 2.0 0.0 PPM Fig. 1 Simulated data with wavelet and SIW baseline fits. x2 - Overall Mean StDev %StDev 1 Wavelet Parametric 1 18624200.0 1 123677.0 / 0.66% Spline Parametric 18710600.0 156170.0 0.83% Wavelet Iterative 18848100.0 118625.0 0.63% Spline Iterative 18905300.0 124918.0 0.66% x2 - Metab Region Wavelet Parametric 29913200.0 199985.0 0.67% Spline Parametric 30031900.0 212750.0 0.71% Wavelet Iterative 30032700.0 192423.0 0.64% 1 Spline Iterative / 30159000.0 1 205769.0 / 0.68% x2 - Lipid Region 1 Wavelet Parametric 1 2803630.0 / 71001.2 1 2.53% 1 Spline Parametric Wavelet Iterative 1 ~~;:Ei / :5403::: I ZE i Spline Iterative 1 3137420.0 / 74224.2 2.37% Table. 1 Averaged results from the Monte Carlo simulation. CONCLUSIONS No significant differences were observed in the accuracy of the fitted metabolite values (typically >5% actual values). The fitted baseline results were comparable for all methods in the metabolite region, where a slowly varying baseline was assumed, but greater variability was observed in fits of the more rapidly changing lipid region, with improved performance obtained by the iterative approach. Fully parametric fitting methods tended to take 3-10 times longer to optimize than the iterative methods. ACKNOWLEDGEMENTS This work was supported by PHS grant ROlAG12119, and F32NS10414 (BJS). REFERENCES 1) SW Provencher, Magn.Res.Med. 30 672-79 (1993) 2) BJ Soher, et.al., Mugn.Res.Med. 40 822-31 (1998) Proc. Intl. Sot. Mag. Reson. Med. 8 (2000) 1871