Biomedical Engineering Reference
In-Depth Information
false-positive (FP) and false negative (FN) voxels and the definition of the fol-
lowing measures:
sensiti
v
ity
=
TP
(
TP
+
FN
)
speci f icity
=
TN
(
FP
+
TN
)
total perf ormance
=
(
TP
+
TN
)
(
TP
+
FP
+
TN
+
FN
)
(2.41)
These measures are very helpful to assess the global performance of a segmen-
tation method such as under-segmentation characterized by a low sensitivity of
over-segmentation characterized by a low specificity. The total performance of
the proposed algorithm stabilized around 98.3% under all noise conditions. The
authors further compared their segmentation performance to morphological op-
erators performance and reported an improvement of sensitivity performance
with the level set method.
2.
A second set of experiments with a database of 18 real brain MRIs of
size (256
×
256
×
176) was performed. Results reported a 94% success ratio of
segmentation convergence (one case failed), requiring on an average 1,000 iter-
ations. Segmentation of individual tissue classes (WM, GM and CSF) required a
coarse approximation of tissue segmentation for class definition and computa-
tion of a priori statistical models.
Limitations:
The proposed segmentation method has a performance limited
by the fact that the SEM algorithm does not guarantee an optimal solution. In
practice, an initial partitioning roughly representative of the inside and outside
distributions of the organs to segment lead to a correct solution. This means that
tissue classes need to be initialized with relatively accurate average intensity
values.
2.4.3.2
Topology Preserving Geometric Deformable Models
for Brain Reconstruction
This research work was published by Han
et al.
in [72].
Method:
The authors proposed a 3D level set segmentation method with
a speed term based on binary flow forces, mean curvature flow and gradient
vector flow. The originality of the method was to focus on the topology of the
evolving front and use the notion of simple points and update the front deforma-
tion only at their locations. Given a set of points defining a 3D surface, a point is
simple if its addition or removal from the object does not change the topology of
