Intensity Segmentation of the Human Brain with Tissue dependent Homogenization

High-precision segmentation of the human cerebral cortex based on
T1-weighted MRI is still a challenging task. When opting to use an
intensity based approach, careful data processing is mandatory to
overcome inaccuracies. They are caused by noise, partial volume
effects and systematic signal intensity variations imposed by
limited homogeneity of the acquisition hardware. We propose an
intensity segmentation which is free from any shape prior. It uses
for the first time alternatively grey (GM) or white matter (WM)
based homogenization. This new tissue dependency was introduced as
the analysis of 60 high resolution MRI datasets revealed
appreciable differences in the axial bias field corrections,
depending if they are based on GM or WM. Homogenization starts with
axial bias correction, a spatially irregular distortion correction
follows and finally a noise reduction is applied. The construction
of the axial bias correction is based on partitions of a depth
histogram. The irregular bias is modelled by Moody Darken radial
basis functions. Noise is eliminated by nonlinear edge preserving
and homogenizing filters. A critical point is the estimation of the
training set for the irregular bias correction in the GM approach.
Because of intensity edges between CSF (cerebro spinal fluid
surrounding the brain and within the ventricles), GM and WM this
estimate shows an acceptable stability. By this supervised approach
a high flexibility and precision for the segmentation of normal and
pathologic brains is gained. The precision of this approach is
shown using the Montreal brain phantom. Real data applications
exemplify the advantage of the GM based approach, compared to the
usual WM homogenization, allowing improved cortex segmentation.