Biomedical Engineering Reference
In-Depth Information
=
−
with
r
p
. Interaction forces often are only computed for a small neighborhood
to reduce computational costs. When using all points the complexity is of order
O
x
n
2
(
)
, while a small neighborhood m with
m
n
can reduce the complexity to
order
O
(
n
·
m
)
.
4.8 Initialization
All the previously presented approaches for segmentation require the placement of
an initial model in the proximity of the desired boundaries. In this section we will
present several possible approaches: manual and landmark-based initialization as
well as automatic methods using the general Hough transform and atlas registration.
4.8.1 Manual Initialization
A simple approach to initialization is to ask the user to place the model manually.
This can be done by using several interaction tools such as the mouse, keyboard or
haptic devices. The manual placement is a time-consuming process and requires an
experienced user, depending on the quality of the images. Nonetheless, this approach
is not usable for large-scale applications. It also reduces the reproducibility, as the
results often depend on the initial model and expertise of the user.
4.8.2 Landmarks
The use of landmarks is an extension of the manual approach. Instead of translating
and scaling the whole model, the user selects a set of strong landmarks in the image
and the initial model is automatically adjusted to obtain the best fit. In this way, the
user input and hence the margin of error can be significantly reduced. Landmarks
should be positioned at significant areas that can be easily detected in the image.
Examples would be areas of high curvature, e.g. a fissure in a bone, or a measurable
location, e.g. the middle of the femural shaft. An initialization of a model using
seeds has been proposed by Neuenschwander et al. [
108
]. The correct scaling and
positioning of the model has to be determined from the seeds. This warping can be
done using the Thin Plate Spine (TPS) transform. If the number of landmarks is very
small and not sufficient for a confident pose estimation, different positions can be
evaluated using a cost function that incorporates a penalty based on image properties.
An example for that can be found in Ref. [
10
].
Using landmarks at key anatomical positions allows the invocation of prior knowl-
edge. Other approaches extend landmarks with inclusion of spatial knowledge. For
