The accuracy and reproducibility of dual-contrast segmentation based on nonparametric feature map analysis have been investigated in a multicomponent gelatin phantom. The root mean square errors in volume ranged from 0.02 cm3 for small volumes to 3.8 cm3 for larger volumes, with a mean error of 0.97 cm3. Average inter- and intraobserver coefficients of variation were found to be < 7% for all compartments. To evaluate the reproducibility of segmentation of clinical image data, volumes of total brain, CSF, and multiple sclerosis (MS) lesions were obtained from five image sets of MS patients. Inter- and intraobserver coefficients of variations were computed for the patient data and were found to be < 5% for brain, 17% for CSF, and 20% for MS lesions. Such variations were found to be reduced by appropriate preprocessing of the images.