Multi-mode Narrow-band Thresholding with Application in Liver Segmentation from Low-contrast CT Images
ABSTRACT Segmentation of liver in CT images is regarded as a challenge in image processing due to low-contrast of datasets, variety of liver shape, and its non-uniform texture; especially for abnormal cases. In this paper, we deal with normal and abnormal datasets as images containing two or more Gaussian components. We threshold a slice in a narrow band of each mode, find liver pixels based on a priori knowledge, prepare a probability map, and threshold the map to find initial liver border. Final boundary of liver is obtained through a few iterations of `Geodesic Active Contour'. The proposed method was tested on 30 normal and 17 abnormal datasets each containing 159-263 slices; acquired from different CT machines. The results for normal and abnormal datasets are completely acceptable, according to the evaluation done by a specialist. However, for severely abnormal datasets, the proposed method is regarded as a promising algorithm for liver segmentation.
- SourceAvailable from: Ron Kimmel
Article: Geodesic Active Contours.[show abstract] [hide abstract]
ABSTRACT: A novel scheme for the detection of object boundaries is presented. The technique is based on active contours evolving in time according to intrinsic geometric measures of the image. The evolving contours naturally split and merge, allowing the simultaneous detection of several objects and both interior and exterior boundaries. The proposed approach is based on the relation between active contours and the computation of geodesics or minimal distance curves. The minimal distance curve lays in a Riemannian space whose metric is defined by the image content. This geodesic approach for object segmentation allows to connect classical "snakes" based on energy minimization and geometric active contours based on the theory of curve evolution. Previous models of geometric active contours are improved, allowing stable boundary detection when their gradients suffer from large variations, including gaps. Formal results concerning existence, uniqueness, stability, and correctness of the evolution a...International Journal of Computer Vision 01/1997; 22:61-79. · 3.62 Impact Factor
- [show abstract] [hide abstract]
ABSTRACT: The aim of this project is computer-based planning of a partial organ resection for three dimensional visualisation of anatomical structures and volumetric estimations. Using quantitative parameters for the estimation of the operation risk, the selection of patients can be further objectified. Another important effect of this project is support for surgical interventions by predefinition of different operation scenarios. All preoperative analysis results are based on tomographic images acquired during the clinical routine. In the end, the integration of planning results into the operation theater as well as into the clinical workflow is part of this project. Surgical strategy is determined by the exact location of the tumor, the respective donor liver part, and the proximity of supporting and draining vessels. Depending on the acquired images, the portal, hepatic, arterial and biliary systems may be analyzed. Therefore, fundamentals for the qualitative and quantitative analysis of images have been developed and implemented since the beginning of the project. The qualitative aspects are, e.g., the three dimensional visualisation of vessel structures, tumors, the security margin and the organ itself. These were evaluated for use in the clinical routine. Preoperatively, the operation strategy is planned by segmenting the important anatomical structures and defining the different vessel systems using a special operation planning software. The visualisation and volumetric results can be visualised and adapted during surgical intervention with the help of a beamer and a touchscreen monitor installed above the patient. In living donor liver transplantation (LDLT) the most important factor is to ensure that the anatomy of the donor patient allows surgical splitting of the liver. Also, a volumetric analysis of the donor liver is of interest to achieving less complicated surgery. Actually, all LDLTs performed in our Surgical Clinic are supported by this operation planning system.Transplantation 01/2004; 78:173-174. · 3.78 Impact Factor
Conference Proceeding: Segmentation of the Liver from Abdominal CT Using Markov Random Field Model and GVF Snakes[show abstract] [hide abstract]
ABSTRACT: Liver segmentation from scans of the abdominal area is an important step in several diagnostic processes. CT scans of the abdominal area contain several organs in close proximity exhibiting similar image characteristics. In this paper, we present preliminary results on an algorithm that uses Markov Random Fields to obtain an initial contour of the liver. Gradient vector fields (GVF) and active contours are used to refine the initial estimate and segment the liver. Tests are reported on 13 clinical cases using a similarity metric that combines area and space.Complex, Intelligent and Software Intensive Systems, 2008. CISIS 2008. International Conference on; 04/2008
Multi-mode Narrow-band Thresholding with Application in
Liver Segmentation from Low-contrast CT Images
Amir H. Foruzana,b, Yen-Wei Chena, Reza A. Zoroofib, Akira Furukawac, Yoshinobu Satod,
aCollege of Information Science and Engineering, Ritsumeikan University, Shiga, Japan.
bControl and Intelligent Processing Center of Excellence, School of Electrical and Computer
Engineering, College of Engineering, University of Tehran, Tehran, Iran.
cDepartment of Radiology, Shiga University of Medical Science, Shiga, Japan.
dDepartment of Radiology, Graduate School of Medicine, Osaka University, Osaka, Japan.
email@example.com, firstname.lastname@example.org, email@example.com
Segmentation of liver in CT images is regarded as a
challenge in image processing due to low-contrast of
datasets, variety of liver shape, and its non-uniform
texture; especially for abnormal cases. In this paper,
we deal with normal and abnormal datasets as images
containing two or more Gaussian components. We
threshold a slice in a narrow band of each mode, find
liver pixels based on a priori knowledge, prepare a
probability map, and threshold the map to find initial
liver border. Final boundary of liver is obtained
through a few iterations of ‘Geodesic Active Contour’.
The proposed method was tested on 30 normal and 17
abnormal datasets each containing 159-263 slices;
acquired from different CT machines. The results for
normal and abnormal datasets are completely
acceptable, according to the evaluation done by a
specialist. However, for severely abnormal datasets,
the proposed method is regarded as a promising
algorithm for liver segmentation.
Development of medical imaging technologies has
made it a necessity to analyze patient datasets before
taking any decision on treatment planning. In case of
liver, its size, volume, and shape; structure of its
vessels; and tumors sizes and locations are
The initial step of liver image analysis is
segmentation. Due to variety of liver’s shape, size, and
location; and low contrast of CT images; liver
segmentation is still considered as a challenge. Many
image processing techniques have been proposed to
develop semi-automatic/automatic, low-level/high-
level, and 2D/3D liver segmentation algorithms. The
techniques which have been employed include
probabilistic atlases , active contours , statistical
shape models , intensity-based , and texture-
based methods . High-level techniques and 3D
approaches are robust but they usually need a training
step. They are sensitive to initialization or registration
step and cannot segment datasets that substantially
differ from the training set. Low-level and 2D
techniques consider variations of liver in different
datasets but are not robust enough and may lead to
over-segmentation or under-segmentation.
In this paper, we propose a novel liver segmentation
algorithm which exploits a 2D technique to consider
variations and uses a 3D technique to establish
robustness for low contrast CT images. In the proposed
method, we first use K-means clustering and a priori
knowledge to find a rough liver boundary in each slice
of the low-contrast CT image. Then the liver
boundaries of all slices are used as the initial surface
input of a 3D geodesic active contour algorithm to find
the accurate liver surface.
2. Narrow band thresholding
The proposed algorithm consists of three steps:
preprocessing, initial boundary extraction, and final
interpolation, smoothing, ROI definition, and initial
slice segmentation. In order to decrease the run-time of
the algorithm and reduce the required memory size, we
extract the region that includes the bounding box of the
trunk in the axial plane, and the bounding box of liver
in the coronal plane. This leads to a reduction of 60%
percent in number of pixels, approximately. The inter-
slice spacing of a dataset is sometimes more than twice
the intra-slice spacing. In such a case, we employ a
cubic interpolator filter to prepare a homogenously-
spaced dataset. Since ribs’ muscles have the same
intensity range as that of liver and they are in contact
with liver, this may lead to over-segmentation in the
result. We extract bones by thresholding and connect
them together by a spline curve to delineate liver and
muscles. Finally, it is needed to segment a slice
manually, which we call it initial slice; in this paper.
This slice should have a large cross-section and
contain the major components of liver intensity. We
adopted the approach used in  to segment liver from
the initial slice up to the first slice, then from the initial
slice down to the last slice. Thus, we do not miss any
separate part of liver.
Initial boundary extraction starts with an analysis of
the initial slice. We
Maximization” algorithm with four Gaussian modes to
estimate mean and standard deviation of major
components which make up the intensity range of
liver. In this analysis, we only consider those
components of the mixture model which have a share
of 5% or more as major liver components. For each
mode, we threshold a slice in narrow region around its
mean value to find liver candidate pixels (Fig. 1(a)).
Let’s intensity range of liver be composed of three
threshold a slice in the range [
The width of this region, i.e. ασ
of standard deviation of the corresponding mode, can
be tuned for high-contrast and low-contrast datasets,
individually. For high-contrast datasets, we select the
fraction in the range 0 [
5 . 0 σ and for low-contrast
datasets, it is in the range
thresholding a slice, we use a priori knowledge of the
previous slice to remove outlier pixels. The remaining
pixels are then clustered by K-means clustering. We
decide to label each cluster as a liver or non-liver
cluster, based on the location of its center (Fig. 1(b)).
If the center of a cluster is inside liver of previous
slice, it is regarded as a liver cluster. We remain all the
members of liver clusters and call them as liver index
pixels. We then threshold the original image in the
whole range of each mode i (
that more pixels are involved. We assign a probability
measure to each pixel based on their
distance () to the index pixels of the object. The
which is a fraction
3 , 2 , 1
, 3 .]
,1 . 0 [
]3 . 0 σ
probability map is shown in
liver region have higher probability. We threshold the
probability map by half of its maximum value to find
initial liver border (Fig. 1(d)).
Finally, the initial liver boundaries of all slices are
used as the initial surface input of a 3D geodesic active
contour algorithm . The final liver boundary is
attained by a few iteration of active contour algorithm.
We tune the number of iterations to a low value so that
the contour is prevented from leakage to nearby
[i μ −
) x (
Fig. 1(c). The pixels in
Fig. 1 Steps of initial liver extraction. (a) Narrow-band
thresholded image, (b) Clustering liver candidate pixels
(Cluster centers are shown in yellow), (c) Probability map,
(d) Initial liver border.
We applied the proposed method to 30 normal
datasets and 17 abnormal datasets, each containing
159-263 images. Datasets belong to Shiga University
of Medical Science and Osaka University with a
resolution of 0.6836 x 0.6836 x 1 mm3 and 0.5859 x
0.5859 x 1.25, respectively.
Preprocessing and active contour parts of the
algorithm were coded in C++ to decrease the run-time
of the algorithm. Initial liver extraction was
implemented in MATLAB 7 to benefit from its rich
image processing toolbox. The platform, on which we
run the algorithm, is an Intel® Core™ 2 Duo with
2GBytes of RAM. The whole algorithm takes 18-22
minutes to segment liver in a dataset.
Segmentation results of sample slices for different
normal and abnormal datasets are shown in Fig. 2
and Fig. 3, respectively. Iso-surface visualizations of
six livers that are segmented by the proposed algorithm
are shown in Fig. 4.
Q ive a s ofu
Table 1-uantitatnalysi the res lts.
In order to evaluate our method quantitatively, we
utilized several measures which are used in MICCAI
2007 Grand Challenge workshop . Quantitative
evaluations for several datasets are shown in Table 1.
Fig. 4 Iso-surface visualization of liver for six
The results of liver segmentation, which are shown
in Fig. 2, correspond to low-contrast datasets. We
included upper slices (Fig. 2(a)), middle slices (Fig.
2(b), (d), (e), (g), (h), Fig. 3(a), (c), (e), (f)), and lower
slices (Fig. 2(c), (f), Fig. 3(b), (d)). Since we use
narrow-band thresholding to select liver pixels and
combine it with a priori knowledge of previously
segmented slice, error of over-segmentation is reduced
Fig. 2 Segmentation resu
lts of sample slices for
. 3 Segmentation results of sample slices for
iff no derent abrmal datasets.
to a large extent. If there is leakage in some slices, due
to k-means clustering, the algorithm can compensate
for it in next slices. Due to small cross-section of liver
in upper and lower slices, initial boundary of liver is
not detected there. However, since we use active
contour algorithm to find final boundary, it improves
the results. For normal datasets, selection of initial
slice has minimum effect on final result. To choose
initial slice, it only suffices to select a slice in which
liver has a large cross-section which is usually found
in the middle of a dataset. To test sensitivity of the
proposed algorithm to the selection of initial slice, we
ran it twice with two different initial slices. The cross-
section of liver in the first initial slice was twice the
first one. However, the results were affected slightly.
Regarding Table 1, segmented volumes have an
absolute relative volume difference of 3.28%.
Maximum surface distance is 30 mm, on average, and
it is usually caused by very narrow regions of liver that
are not detected by the algorithm. Another source of
error is segmentation of Inferior Vena Cava (IVC).
Physicians regard it as part of liver only in slices where
is completely enclosed by liver. We need a rule-
based algorithm to segment IVC separately and decide
on whether it should be attached to the result or not.
As can be seen in Fig. 3, the algorithm has
acceptable results for abnormal datasets. If liver’s
texture changes severely for different slices of an
abnormal dataset, the algorithm will lose liver in these
slices since it cannot detect available modes of
intensity correctly. We need to consider variations of
es. Fig. 4 shows that the proposed algorithm can
segment liver with arbitrary shapes and it does not
need a training step.
We also applied our method to high-contrast
datasets in which we need a different set of parameters,
with respect to low-contrast datasets. The switching
mechanism between two sets of parameters is done by
user. We intend to use an automatic mechanism to
de datasets based on the contrast of images. Since
we find final boundary of liver through a 3D
algorithm, it exploits robustness of 3D approache
By starting the task of segmentation from
slice in which liver has a single cross-se
n miss any separate part of liver in next slices.
to have good results for such
ction, we do
5. Conclusion and future works
In this paper, we presented a new method for
segmentation of liver from low-contrast CT datasets.
We proved the ability of the method by applying it on
low-contrast CT datasets. The results of segmentation
for normal datasets are good and those for abnormal
rent imaging machines. Also, we decide to
heir contrast into two or
propriate set of parameters
more groups and use an ap
This work was supported in part by the Grand-in
Aid for Scientific Research from the Japanese Ministry
for Education, Science, C
Grand No. 21300070
ulture and Sports under the
and in part by the Research fund
n Global Innovation Research
Planning for Living Liver Donation,” 20th ISPRS Congress,
Istanbul 2004, International Archives of Photogrammetry
and Remote Sensing, XXXV(B), 2004, pp. 291-295.
. Bland, C. Meyer. ”Construction of an
abdominal probabilistic atlas and its application in
segmentation,” IEEE Transaction on Medical Imaging, 2003,
Markov Random Field Model and GVF Snakes,”
Proceedings of the 2008 Intl. Conf. on Complex, Intelligent
and Software Intensive Systems, 2008, 00:293-298.
e, C. Koehl, et al., ”Fully automatic anatomical,
pathological, and functional segmentation from CT scans for
hepatic surgery,” Computer Aided Surgery, 2001, 6(3):131-
C. Güzeliş, “Patient oriented and robust automatic liver
segmentation for pre-evaluation of liver transplantation,”
Computers in Biology and Medicine, 2008, 38(7):765-784.
rst J. “A Hybrid
MICCAI Workshop 3D
datasets are promising. In future, we try to expand the
proposed method to include more abnormal datasets
grade liver datasets based on t
 H.P. Meinzer, P. Schemmer, M. Schobinger, M. Nolden,
T. Heimann, B. Yalcin, and et al., “Computer-based
 H. Park, P
 R. S. Alomari, S. Kompalli, V. Chaudhary.
“Segmentation of the Liver from Abdominal C
 L. Soler, H. Delingette, G. Malandain, J. Montagnat, N.
 M. Selver, A. Kocaoğlu, G. Demir, H. Doğan, O. Dicle
 Susomboon R, Raicu DS, and Fu
Approach for Liver Segmentation.”
Segmentation in the Clinic: A Grand Challenge, T. Heimann,
M. Styner, B. van Ginneken (eds.), 2007.
 V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active
contours,” International Journal on Computer Vision, 1997,
 Ginneken BV, Heimann T, Styner MA (2007) 3D
segmentation in the clinic: a grand challenge. In:Workshop
on 3D segmentation in the clinic, MICCAI 2007, pp 7–15.