An accurate segmentation of white matter, gray matter and cerebrospinal fluid (CSF)
in MR images of the brain is key to understanding important brain diseases. We present
a new system for segmentation of MR images of the brain, based on a novel modification
of the watershed transform. Our proposed improvement is to substitute the single contour
detection function (usually the gradient) of the original watershed transform with
a set of functions especially tailored for the detection of each structure in the
image. In this paper, these functions are based on a previous probability calculation,
using normal distributions and a Markov Random Field. To improve the detection of
the sulci, where the partial volume effect often masks the presence of CSF, the probability
values for gray matter and CSF are modified using the absolute value of the distance
to the white matter and the ridgeness of this distance. We also propose a novel way
to initialize the watershed transform by using a probabilistic atlas: in this way,
no user interaction is needed. Validation experiments indicate an accurate segmentation
of the interesting structures.
