Biomedical Engineering Reference
In-Depth Information
While manual segmentation of the complicated cortex is time and labor inten-
sive and dependent upon the individual user, semi- or fully automatic segmentation
is indispensable. There is a growing number of efforts in this arena. Wells et al.
[8] used a non-parametric Pazen window to model brain images, and corrected
RF inhomogeneities using an EM algorithm. Kapur [92] extended Wells' work by
incorporating spatial information embedded in the raw data in conjunction with
an EM-MF estimator to intensively compute inhomogeneities in the RF coil and
aid in classification of pixels. Hall et al. [25] employed the FCM in classification
of brain images. Xu et al. [93] proposed an Adaptive FCM (ACFM), based on
the FCM, for the purpose of correcting the bias field. Bomans et al. [94] ap-
plied morphological filters, along with a Marr-Hildreth operator, for 3D cortical
segmentation and reconstruction.
In the meantime, deformable models, and level set methods in particular,
have become widely recognized for their great potency in segmenting, tracking,
and matching anatomic structures by information derived from image data, as well
as prior knowledge about the location, size, and shape of structures. Leventhon
[95,96] incorporated a priori shape information into the level set framework as the
shape force evolving the curve to match the expected shape perfectly. Han et al.
[97] proposed topologically preserving the level set, which retains initial topology
during evolvution of the curve. Zeng et al. [59] proposed a novel coupled level
set and applied it to modeling of the cortical structure. It should be mentioned
here that our work is somewhat close to Zeng's work, but the major distinctions
between them will be discussed in later sections.
In this section we present a novel approach to cerebral cortex segmentation.
We formulate the segmentation problem within the level set framework, the zero
level set surface of which is driven by image-derived information, and the a priori
knowledge of cortex structures as well. Its efficacy is demonstrated with experi-
ments where this formulation addresses the cortex segmentation problem robustly
and accurately.
The outline of this section is as follows. In Section 4.2 we provide a preview of
our framework, followed by a detailed description of its components and their in-
tegration into the final variational formula, which propagates the surface within the
WM and localizes it at WM/GM boundary. The WM/GM surface further evolves
into the GM and halts at the GM/CSF boundary. In Section 4.3 we demonstrate
our approach with simulated as well as real MR brain data, and provide visual and
quantitative results. Finally, we draw our conclusions in Section 4.4.
5.2. Our Approach
In this subsection we develop a two-stage semiautomatic cortex segmentation
system, as depicted in Figure 10. Since skulls are part of the brain, we first remove
them with the aid of MRIcro software, which is a freeware used for visualization
and processing of medical images. We next refine the results through manual
