![](bg6.png)
The method attained a good performance in 17 of
the 20 exams, in which the overall score is above 65.
Figure 1 shows the result obtained in the best case,
with overall score of 82.05. It is possible to observe
that the liver boundaries are accurately defined.
In 3 exams, though, the results contain some
significant errors that can be verified visually. The
exam with the lowest score among all exams tested
has an overall score of 29.57. In this exam the liver
has a huge nodule, and it causes a leak of the
segmented region towards adjacent darker
structures. Considering the size of the nodule, the
result is reasonable, though.
On another exam with low score (46.18) it is
possible to observe a single major error, caused by a
peripheral nodule not classified as liver. This is
explained by the heuristic adopted, that considers a
single Gaussian curve to model liver tissue. Once the
nodule in this exam appears much darker than the
liver parenchyma, its voxel intensities lie outside the
range [TL,TH] defined by the Gaussian fit. As the
nodule is peripheral, it wasn’t possible to correct this
error with morphological fill holes, and therefore the
nodule region was not included in the final result.
Our results can be easily compared with many
other approaches, since the data and evaluation
metrics were obtained from the website of the liver
segmentation competition held in the Sliver07
conference, and the results of other approaches are
also available there. Thus, this comparison with
other works is straightforward once one visits the
conference’s website. If compared with other
automatic and semi-automatic methods, our method
has a good performance being ranked among the top
5 score.
6 CONCLUSIONS
We have presented a method to segment the liver
based on a level sets approach, using an evolutionary
method to estimate its optimal parameters. These
parameters were coded into genes of the individuals
of a GA, and the fitness evaluation was defined to
measure the similarity between a user defined
reference and the segmentation result.
Trough all the experiments it was possible to
verify the potential of the presented methodology.
The use of level sets, which is a consolidate
alternative to segment medical images, achieved
good performances in the tested exams, and the use
of GA to estimate its optimal parameters produced
robust parameters.
The method has, though, some limitations. It
presented some low performances in the presence of
peripheral nodules and veins, and also when nodules
with volume similar to the liver parenchyma were
observed. These cases were presented in details in
section 5.
It is important to notice that the method can be
applied to segment other organs beside the liver,
especially considering the ones roughly
homogeneous. In this case the GA would estimate
other parameters based on the input reference of the
organ to be segmented.
Some suggestions for further research would be a
better modelling to build the speed image
considering also the information of liver internal
structures, such as vessels and nodules. Another
possibility would be use the advection term to
suppress or reinforce some specific barriers, which
could be used to avoid leaking and also enable the
inclusion of peripheral nodules and veins in the final
result.
REFERENCES
Davis L. Handbook of Genetic Algorithms. Van Nostrand
Reinhold Company, New York, 1990.
Fujimoto H., Gu L., and Kaneko T., “Recognition of
abdominal organs using 3D mathematical
morphology,” Trans. Inst. Electron. Inf. Commun.
Eng. D-II, no. 5, pp. 843-850, May 2001.
Heimann T.; van Ginneken B.; Styner M.. "3D
Segmentation in the Clinic: A Grand Challenge",
(Eds.): 3D Segmentation in the Clinic: A Grand
Challenge, pp. 7-15, 2007.
Malladi, R. ,Sethian, J.A., Vemuri, B.C.. Shape modeling
with front propagation: a level set approach. IEEE
rans. Pattern Anal. Mach. Intell. 17 (2) 158–175, 1995.
Michalewicz Z (1994) Genetic Algorithms + Data
Structures = Evolution Pro-grams. Springer-Verlag,
Berlin Heidelberg New York
Lamecker H., Zachow S., Haberl H., Stiller M.. Medical
Applications for Statistical 3D Shape Models. Proc.
Computer Aided Surgery Around the Head, volume 17
of Fortschritt- Berichte VDI, p. 61, 2005.
Osher S., Sethian J. Fronts propagating with curvature-
dependent speed: algorithms based on Hamilton-
Jacobi formulations. J. Comput. Phys. 79 (1998) 12–
49.
Yoo T.S., Ackerman M. J., Lorensen W. E., Schroeder W.,
Chalana V., Aylward S., Metaxes D., Whitaker R..
Engineering and Algorithm Design for an Image
Processing API: A Technical Report on ITK - The
Insight Toolkit. In Proc. of Medicine Meets Virtual
Reality, J. Westwood, ed., IOS Press Amsterdam pp
586-592 (2002).
LIVER SEGMENTATION USING LEVEL SETS AND GENETIC ALGORITHMS
159