To measure the accuracy of single-voxel, image-guided proton MR spectroscopy in distinguishing normal from abnormal brain tissue and neoplastic from nonneoplastic brain disease. MR spectroscopy was performed at 0.5 T with the point-resolved spectroscopic pulse sequence and conventional postprocessing techniques. Subjects consisted of a consecutive series of patients with suspected brain neoplasms or recurrent neoplasia and 10 healthy adult volunteers. Fifty-five lesions in 53 patients with subsequently verified final diagnoses were included. Spectra were interpreted qualitatively by visual inspection by nonblinded readers (prospectively) with the benefit of prior clinical data and imaging studies, and by blinded readers (retrospectively). The nonblinded readers interpreted the spectra as diagnostic or not, and, if diagnostic, as neoplastic or nonneoplastic. The blinded readers classified the spectra as diagnostic or not, and, if diagnostic, as normal or abnormal and as neoplastic or nonneoplastic (when abnormal). The sensitivity, specificity, positive and negative predictive values, and accuracy were calculated from blinded and nonblinded MR spectroscopy interpretations. A receiver operator characteristic (ROC) curve analysis was performed on blinded MR spectroscopy interpretations. The diagnostic accuracy averaged across four blinded readers in differentiating patients from control subjects was .96, while the area under the aggregate (pooled interpretations) ROC curve approached unity. Accuracy in the nonblinded and blinded discrimination of neoplastic from nonneoplastic disease was .96 and .83, respectively. The area under the aggregate ROC curve in the blinded discrimination of neoplasm from nonneoplasm was .89. Image-guided proton spectra obtained at 0.5 T from patients with suspected neoplasia can be distinguished from spectra in healthy control subjects, and neoplastic spectra can be distinguished from nonneoplastic spectra with a high degree of diagnostic accuracy.