The histopathological characterization of brain tumors requires the use of biopsy, which is an invasive surgical procedure. However, a biopsy does not always show the real grade of the tumor because of tumor heterogeneity. Many studies have been done to assess the utility of in vivo proton magnetic resonance spectroscopic imaging ( 1 H MRSI) in noninvasive diagnosis of brain tumors [1][2][3][4][5][6][7][8][9][10][11][12][13] . The earliest studies [2][3][4][5]8 used single-voxel spectroscopy (acquisition of a spectrum for one volume of interest inside the tumor) and classification into histopathological groups based on the study of peak ratios. More recently, pattern recognition techniques have been used to improve the classification of the different brain tumor types 6,7,9,10,12,13 . These techniques allow an automatic classification of spectra based on an analysis of the whole spectrum after digitization or of a profile extracted from selected peak intensities 14,15 . Promising results have been obtained, with accuracy rates of 55-100%. However, individual variability, mainly due to the heterogeneity of the tissue within a voxel, results in a substantial overlap between different brain tumor types 2 and between high-and low-grade glial tumors 8 .Another approach to improving brain tumor characterization is based on 1 H MRSI in which spectra from all voxels of a selected slice are acquired simultaneously. 1 H MRSI shows that metabolites are distributed heterogeneously in the tumor 1,11,16 , but it is difficult to extract information about the histopathology of each voxel and about the histopathological heterogeneity of the tumor. A spectroscopic examination leads to a large amount of information: 64-100 spectra containing up to six peaks (Fig. 1), attributed to cholinecontaining compounds (3.2 parts per million (p.p.m.)), creatinecontaining compounds (3.0 p.p.m.), N-acetyl-aspartate (NAA; 2.0 p.p.m.), alanine (1.4 p.p.m.), lactate and/or lipids (1.3 p.p.m.) and lipids (0.9 p.p.m.). Several images of metabolite distribution (metabolite maps) can be obtained for each patient (Fig. 1). It is very difficult to extract pertinent information from visual inspection of these metabolite maps. The analysis might be limited to one or two voxels of the spectroscopic images, but the difficulty lies in deciding which ones are most representative of the tumor. One region of interest (ROI) might be defined in the tumor, with comparison of the averaged metabolite ratio obtained over this ROI (ref. 7), but then the information about tumor heterogeneity is lost. The main difficulty is combining the information contained in six metabolite maps into a single image giving voxel-by-voxel information about the tumor.Here we present a new approach for extracting information about brain tumor types and grades from 1 H MRSI data. We sought to design a tool that could be used routinely and easily by radiologists for assessing the potential of 1 H MRSI for brain tumor diagnosis, studying heterogeneity, stereotactic biopsy guidance and thera...