Quantitative ‘H Spectroscopic Imaging of Human Brain at 4.1 T Using Image Segmentation zyx Hoby P. Hetherington, Jullie W. Pan, Graeme F. Mason, Dorothy Adams, Michael J. Vaughn, Donald B. Twieg, Gerald M. Pohost zyxw Metabolic differences in the content of N-acetylaspartate (NAA), creatine (CR), and choline (CH) in cerebral gray and white matter can complicate the interpretationof ‘H spectro- scopic images. To account for these variations, the gray- and white-matter content of each voxel must be known. To pro- vide these data, a T,-based image segmentation scheme was implemented at 4.1 T. The tissue composition of each voxel was determined using the point-spread function of the spec- troscopic imaging acquisition and the segmented anatomical image. Pure gray- and white-matter values for CR/NAA and CH/NAA, and the content of CR, CH, and NAA, were deter- mined using a linear-regression analysis of 984 voxels ac- quired from 10 subjects using white-matter CR as an internal standard. This information was used to establish means and confidence intervals for CRINAA and CH/NAA from a voxel of arbitrary tissue composition. Using a single-tailed zyxwvutsr t test, the extent and locations of the metabolic abnormalities zyxwvuts (P zyxwvuts < 0.05) in a patient with multiple sclerosis were identified. Key words: ’H spectroscopic imaging; image segmentation; quantitation; multiple sclerosis. INTRODUCTION zyxwvuts ‘H spectroscopic imaging (SI) of cerebral N-acetylaspar- tate (NAA), creatine (CR), and choline (CH) has been applied to a variety of clinical states including multiple sclerosis (MS) (I), cancer (z), stroke (3), and epilepsy (4). Analysis of these data is usually performed by either a qualitative evaluation of the metabolite images or a quan- titative evaluation of selected voxels from regions of in- terest. Unfortunately, because of the differences in gray- matter (GM) and white-matter (GM) concentrations of NAA, CR, and CH (5-lo), significant variations in metab- olite ratios can be seen due solely to differences in the mixture of GM and WM in the voxel, thereby limiting the interpretation of the metabolite ratios. When screening for neuronal loss using the NAA content (11) is con- ducted, inclusions of varying amounts of cerebral spinal MRM 3621-29 (1996) From the Departments of Medicine (H.P.H., G.F.M., G.M.P.), Neurology (J.W.P.), Biomedical Engineering (D.B.T.), University of Alabama at Birming- ham, Birgmingham, Alabama; and the Department of Neuroradiology (D.A.), Columbia University, New York, New York. Address correspondence to: Hoby P. Hetherington, Ph.D., Center for Nu- clear Imaging Research, University of Alabama at Birmingham, 828 8th Court South, Birmingham, AL 35294. Received August 10, 1995; revised February 8, 1996; accepted February 12, 1996. Copyright zyxwvutsrqponm 0 1996 by Williams & Wilkins All rights of reproduction in any form reserved. 0740-3194/96 $3.00 fluid (CSF) can mimic the presence of neuronal loss or damage. Previous single-voxel studies have overcome these problems by a careful choice of the volume of interest to minimize variations in GM and WM content (5, 6) or applied corrections for inclusion of CSF (9, 12). However, when the exact site or extent of the lesion is unknown, the SI approach retains significant advantages. Therefore, to overcome the limitations of partial-volume effects and fully utilize the information content available in a SI data set, the composition of each voxel with respect to GM, WM, and CSF must be known (13-17). The decomposition of a brain image into its relevant major subcomponents, GM, WM, and CSF has been an area of extensive investigation (18-21). These ap- proaches have included the use of zyxw T,, T,, and proton- density-weighted images to provide a quantitative sepa- ration of GM and WM. Recently, Pan et al. (22) and Ugurbil et al. (23) demonstrated that the Tl contrast at 4 T is sufficient to provide an excellent delineation of GM, WM, and CSF in inversion-recovery-based images, sug- gesting that image segmentation could be achieved using the TI. However, the quantitative measurement of tissue T,s using conventional multiple-image acquisitions with varying inversion recovery times (TIRs) is very time con- suming. Recently, a variety of methods for the rapid acquisition of Tl images have been reported (23-26). These methods decrease the required acquisition time by sampling the magnetization at multiple points along a single recovery curve using low flip angles. Applying a similar approach, we have used a 10-point three-slice measurement to acquire T, images of the hu- man brain at 4.1 T. Using the calculated Tl data, we have classified each pixel within the images as either GM, WM, CSF, or arising from extracerebral (EC) tissues. With these assignments along with the point-spread function of the ‘H SI acquisition (lo), we have calculated the tissue composition of contiguous blocks of 96-100 vox- els from SI studies in 10 healthy volunteers. To evaluate the effectiveness of this procedure, we correlated the CRINAA ratio with the amount of GM in each voxel. We calculated the content values of CH, CR, and NAA from the linear-regression determined values for “pure” GM and WM using WM CR as an internal concentration stan- dard and using corrections for TI, zyx T2, and the frequency dependence of the sequence. These results are compared with previous studies using single-voxel measurements at 1.5 and 2.0 T and a recent SI study with image seg- mentation (13). Finally, as a demonstration of the utility of this method, we used the pooled data from the 10 21