Biomedical Engineering Reference
In-Depth Information
4.
DEFORMABLE REGISTRATION OF BRAIN ATLASES TO BRAIN
TUMOR IMAGES
4.1. Introduction and Background
Today, radiologists, neurosurgeons, and neurooncologists use 3Dbrain images
of brain tumor patients for diagnosis, therapy, and surgery planning. In addition
to this everyday use of brain tumor images at clinical centers, the deformable
registration of these images into a common stereotactic space has the potential
of making a significant positive impact on brain cancer research and therapeutic
outcome. First, this deformable registration will make it possible to apply the
morphometric analysis methods described above in this chapter to brain tumor im-
ages. Perhaps more importantly, solving this deformable registration problemwill
allow the pooling of data from a large number of patients into a common stereo-
tactic space, which will make possible the construction of statistical atlases that
are based on collective morphological, functional, and pathological information.
These atlases will be useful in planning neurosurgical operations and therapeutic
approaches that deal with brain tumors by statistically linking the functional and
structural information provided by multimodality radiological images to variables
such as tumor size, grade, progression, therapeutic approach, and outcome. The
registration of such atlases to a patient's tumor-bearing images will augment the
images with the rich information available in the atlases and therefore will act as
tools for optimal planning of tumor treatment.
Since current anatomical brain atlases are based on images of normal human
subjects, tumors are absent from these atlases. The direct application of currently
available deformable registration methods to adapt the images of a neuroanatomy
atlas to a patient's tumor-bearing images is therefore inundated by inaccuracies
near the tumor due to the substantial dissimilarity between the two images. These
dissimilarities arise from topological differences between the atlas and the patient's
images, severe deformation in the vicinity of the tumor, tumor infiltration, tissue
death and resorption, and the confounding effects of edema.
To account for topological differences between the atlas and the patient's
images, Dawant et al. [77] suggested an approach that introduces a tumor “seed”
in the brain atlas, which is subsequently grown to match the patient's tumor via the
used deformable image registration method. Although this approach is successful
in producing qualitatively acceptable results, the seed deformation is strongly
dependent on the parameters of the image-matching method, as well as the seed
location and size, which are only chosen in an approximate manner. To overcome
the problems of this approach, Bach Cuadra et al. [78] suggested the use of a
simplified, radially symmetric model of tumor growth coupled with the image-
matching forces. In spite of this improvement, the approach in [78] was still not
able to achieve acceptable atlas registration accuracy for patients with large tumors
and substantial brain tissue deformation. Such inaccuracies can be traced back
