Magnetic resonance (MR) imaging now allows the qualitative and quantitative assessment of the human brain in vivo. As MR imaging resolution has improved, precise measurement of small brain structures has become possible. Methods of measuring brain regions from MR images include both manual and semiautomated methods. Despite the development of numerous volumetric methods, there have been only limited attempts so far to evaluate the accuracy and reproducibility of these methods. In this study we used phantoms to assess the accuracy of the segmentation process. Our results with simple and complex phantoms indicate an error of 3-5% using either manual or semiautomated techniques. We subsequently used manual and semiautomated volumetric methodologies to study human brain structures in vivo in five normal subjects. Supervised segmentation is a semiautomated method that accomplishes the division of MR images into several tissue types based on differences in signal intensity. This technique requires the operator to manually identify points on the MR images that characterize each tissue type, a process known as seeding. However, the use of supervised segmentation to assess the volumes of gray and white matter is subject to pitfalls. Inhomogeneities of the radiofrequency or magnetic fields can result in misclassification of tissue points during the tissue seeding process, limiting the accuracy and reliability of the segmentation process. We used a structured seeding protocol that allowed for field inhomogeneity that produced reduced variation in measured tissue volumes. We used repeated segmentations to assess intra- and inter-rater reliability, and were able to measure small and large regions of interest with a small degree of variation. In addition, we demonstrated that measurements are reproducible with repeat MR acquisitions, with minimal interscan variability. Segmentation methods can accurately and reliably measure subtle morphometric changes, and will prove a boon to the study of neuropsychiatric disorders.