Use of Voigt Lineshape for Quantification of zy in zy Vivo ‘H Spectra zyxwv Ian Marshall, John Higinbotham, Stephen Bruce, Andreas Freise zyx Quantification of NMR visible metabolites by spectral model- ing usually assumes a Lorentzian or Gaussian lineshape, de- spite the fact that experimental lineshapes are neither. To minimize systematic fitting errors, a mixed Lorentzian-Gaus- sian (Voigt) lineshape model was developed. When tested with synthetic FIDs, the Voigt lineshape model gave more accurate results (maximum error 2%) than either Lorentzian(maximum error 20%) or Gaussian models (maximum error 12%). The three lineshape models gave substantially different peak ar- eas in an zyxwvutsrqp in vitm experiment, with the Voigt model having a much lower 3 (2.1 compared with 5.2 for the Lorentzian model and 6.2 for the Gaussian model). In a group of 10 healthy volunteers, fitting of ’H spectra from cerebral white matter gave significantly different peak areas between the methods. Even when area ratios were taken, the Lorentzian model gave higher values (+5% for “Vcholine and +2% for Nucreatine) than the Voigt lineshape model, whereas the Gaussian model gave lower values (-2% and zyxwvut -1v0, respec- tively). Key words: NMR spectroscopy; quantification; modeling; fit- ting. INTRODUCTION Accurate quantification of zyxwvutsrq in vivo proton NMR spectra is an important aspect of the study of many brain disorders (1, 2). Integration of metabolite peak areas between se- lected frequency limits is long-established, and suitable software is available on many commercial imaging sys- tems. It is not particularly accurate, however, in the presence of noise and overlapping peaks as encountered in experimental spectra and is somewhat operator-de- pendent. Here we consider numerical modeling tech- niques for fitting the metabolite data to assumed FID or lineshape models (3-8). Most of these methods assume that the metabolites have exponentially decaying signals, which give rise to Lorentzian lineshapes in the frequency domain. Experimental data, however, whether in vitro or in vivo, are not purely Lorentzian. Even though metabo- lite signals may be intrinsically monoexponential, imper- MRM 32651-857 (1997) zyxwvutsrqpo From the Department of Medical Physics and Medical Engineering, Univer- sity of Edinburgh, and Department of Applied Chemical and Physical Sci- ences, Napier University, Edinburgh (J.H.. S.B.. A.F.). Address correspondence to: Ian Marshall, Ph.D.. MRI Unit, Westem General Hospital, Edinburgh EH4 ZXU, United Kingdom. Received June 24, 1996; revised October 17. 1996 accepted October 28. 1996. This work was funded in part by the U.K. Medical Research Council, under the Clinical Research Initiative in Neurosciences. The HLSVD software was funded by the European Community “Human Capital and MobilityMet- works” project. Present address (A.F.): Department of Physics, University of Hannover, Germany. Copyright 0 1997 by Williams 8. Wilkins All rights of reproduction in any form resewed. 0740-31 94/97 $3.00 fect shimming and susceptibility variations cause a spread of resonant frequencies within the spectroscopic volume of interest (VOI). Pragmatically, the probability distribution function of the individual lineshapes can be approximated by a Gaussian function, leading to a Gaus- sian broadening (convolution) of the underlying line- shape. Thus the spectral peaks to be modeled are Lorent- zian with Gaussian line-broadening, resulting in Voigt lineshapes (9). The relative linewidths of the Lorentzian and Gaussian components (the Voigt parameter) are not known a priori. In contrast, established spectral model- ing techniques use fixed lineshapes, which are usually Lorentzian, although Gaussian lineshapes have occasion- ally been used (7). Voigt lineshapes have been discussed in the context of infrared (10) and electron spin reso- nance (11) spectroscopy, but have rarely been used for NMR (12). The popular time-domain method HLSVD (5) uses exponentially damped sinusoids (Lorentzian line- shapes). VARPRO (4), another time-domain method, can in principle be used with an arbitrary lineshape function, but in practice is normally used to fit Lorentzians or Gaussians. Concerned that the usually fixed lineshape methods introduce bias into metabolite quantification, we devel- oped a frequency-domain Voigt model that combines Lorentzian and Gaussian lineshapes (13). A systematic survey of the quantification errors caused by using inap- propriate lineshapes was conducted by comparing the performance of the Voigt model on synthetic ‘H NMR data with Lorentzian-only and Gaussian-only models. The same data was also evaluated with the HLSVD method. The three lineshape models and HLSVD were then applied to in vitro and in vivo ‘H spectroscopic data. METHOD All analysis was conducted on a SPARC-10 workstation (Sun Microsystems, Mountain View, CA) running Unix. The HLSVD fitting software was acquired from the au- thors, while software for the generation of synthetic data and frequency-domain modeling was written in-house by using the “C” language. Simulations Synthetic FIDs were generated consisting of 4096 data points at 1 ms intervals. Metabolites were simulated with chemical shifts of 3.2,3.0, and 2.0 ppm, and time-domain amplitudes of 1.0, 1.0, and 2.5 units, respectively. Each metabolite was given an intrinsic T2 of 300 ms, and zero phase. The full width at half maximum (FWHM) of the corresponding Lorentzian lineshape is given by 1/( TTJ = 1.1 Hz, or 0.017 ppm at a field strength of 1.5 T. This model corresponds to reported values for choline, creat- ine, and NAA in healthy subjects (14). The FIDs were 651