3D Finite Element Model of a Female Pelvis Reconstructed from
CT Images
Christopher R. Dance, Rudy J. A. Lapeer, Richard W.
Prager
Cambridge University Engineering Department
ABSTRACT
This paper discusses the generation of a finite element model
of a
female pelvis from CT images, obtained from the Visible
Human
Project. The pelvis model is an essential part of the simulation
of a
childbirth process which is the ultimate objective of this
research.
The images are segmented using a semi-automatic approach
based on
active contours. A novel algorithm involving Delaunay
triangulation
is used for reconstruction. This algorithm is more efficient
than
voxel reconstruction methods and can cope with arbitrary planar
data
not necessarily from parallel planes. A mesh is created which
allows
modelling of the deformation of the pelvis and fetal head
during
delivery.
SEGMENTATION
A 3D simulation of the childbirth process, ready for use in a
clinical environment, should be designed in a way that makes it
instantly available after gathering the necessary images. This
objective calls for an automated approach. Automatic
reconstruction from contours does not pose a problem as will be
shown in the next section. However fully automated segmentation
of CT or MR images is one of the difficult issues in medical
image processing.
A large variety of segmentation techniques exist which can be
roughly divided into pixel classification, region-based and
edge-based segmentation techniques [4].
In this work, the Canny operator was initially applied to
segment the
image. Then gradient thresholding was used to create closed
contours
surrounding the regions of interest. Although the Canny
approach
guarantees good detection and localisation and uniqueness of
response
to edges [3, 4], results produced from medical
images are
unsatisfactory because:
- double edges occur due to inhomogeneity in the bone tissue;
- edges are not fully connected since the gradient at the bone
boundaries varies;
- false edges occur because other tissue boundaries have
gradients in the same range as those of the bone. This noise
cannot be removed by gradient thresholding.
This led us to decide on a semi-automatic approach, however
restricting human interaction to a minimum. Active contours are
used, with an external energy equal to the reciprocal of the
magnitude of the gradient of a Gaussian convolved with the image.
[Next]