Biomedical Engineering Reference
In-Depth Information
information [22, 23] to make the segmentation process more robust to poor
edge definition. When dealing with weak boundaries the most radical solution
to leaking problems is to remove the expansion term at the cost of requiring an
initialization close to the final solution [24]. An alternative to this approach was
proposed by Jin
et al.
[25] initially keeping the expansion term for pushing the
model and turning it off as it approaches the object boundary. Detection of the
boundary location was performed using a homogeneity map derived from scale-
based fuzzy connectivity [26]. A more recent effort to address the problem of
segmentation of an object with missing boundaries was presented by Sarti
et al.
in [27] introducing a new geometric model for subjective surfaces. Starting from
a reference point inside the object to segment, the “point of view”, a geometric
deformable model is evolved with mean curvature flow and image-derived speed
terms until a piecewise constant solution is reached. This piecewise constant
solution is the subjective surface defined by the segmentation process that is
flat inside the object and has boundaries defined by geodesic curves. The au-
thors also introduced the notion of “modal” contours which are contours that
are perceived in the visual context and “amodal” contours which are associated
with partially occluded objects. Segmentation of amodal contours can be per-
formed with their subjective surface framework through iterations of edge-map
computation and contour extraction. The authors produced very nice illustra-
tions of the performance of their subjective surface segmentation on three-
dimensional ultrasound data with a fetal echogram, recovering the shape of the
fetus.
All the level set segmentation methods presented above are based on image
gradient intensity making them prone to leaking problems in areas with low
contrast. A second problem related to the use of the image gradient as the
only image-derived speed term is that it makes the segmentation process very
sensitive to the initial position of the level set function as the model is prone
to converge to false edges that correspond to local minima of the functional.
Medical images typically suffer from insufficient and spurious edges inherent to
physics of acquisition and machine noise from different modalities.
Two approaches can be followed to address these limitations. The first ap-
proach is to fuse regularizer terms in the speed function as reviewed in [9]. A
second approach is to reformulate the problem in terms of region-based seg-
mentation methods derived from the Mumford-Shah functional implemented in
