Ce. Byrum et al., ACCURACY AND REPRODUCIBILITY OF BRAIN AND TISSUE VOLUMES USING A MAGNETIC-RESONANCE SEGMENTATION METHOD, PSYCHIATRY RESEARCH-NEUROIMAGING, 67(3), 1996, pp. 215-234
Magnetic resonance (MR) imaging now allows the qualitative and quantit
ative 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 b
oth manual and semiautomated methods. Despite the development of numer
ous volumetric methods, there have been only limited attempts so far t
o evaluate the accuracy and reproducibility of these methods. In this
study we used phantoms to assess the accuracy of the segmentation proc
ess. 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 segmenta
tion is a semiautomated method that accomplishes the division of MR im
ages into several tissue types based on differences in signal intensit
y. 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 vol
umes of gray and white matter is subject to pitfalls. Inhomogeneities
of the radiofrequency or magnetic fields can result in misclassificati
on of tissue points during the tissue seeding process, limiting the ac
curacy and reliability of the segmentation process. We used a structur
ed seeding protocol that allowed for field inhomogeneity that produced
reduced variation in measured tissue volumes. We used repeated segmen
tations to assess intra- and inter-rater reliability, and were able to
measure small and large regions of interest with a small degree of va
riation. In addition, we demonstrated that measurements are reproducib
le with repeat MR acquisitions, with minimal interscan variability. Se
gmentation methods can accurately and reliably measure subtle morphome
tric changes, and will prove a boon to the study of neuropsychiatric d
isorders.