Automatic Colon Segmentation Using Isolated-Connected Threshold
Ku-Yaw Chang*, Hao-Han Zhang*, Shao-Jer Chen#,¶, Lih-Shyang Chen§, Jia-Hong Chen*
*
Department of Computer Science and Information Engineering, Da-Yeh University, Changhua, Taiwan, R.O.C.
e-mail: canseco@mail.dyu.edu.tw, gama02174609@hotmail.com, xzone0826@gmail.com
#
Department of Radiology, Buddhist Dalin Tzu Chi General Hospital, Chiayi, Taiwan, R.O.C.
¶
School of Medicine, Buddhist Tzu Chi University, Hualien, Taiwan, R.O.C.
e-mail: shaojer.chen@msa.hinet.net
§
Department of Electrical Engineering, National Cheng-Kung University, Tainan, Taiwan, R.O.C.
e-mail: chens@mail.ncku.edu.tw
Abstract— Virtual colonoscopy (VC) is a safe and fast medical
imaging procedure to screen the colon for polyps. And it has
become very popular recently. Colon segmentation is a
necessary and important step of such an examination
procedure. In this paper, an automatic colon segmentation
method is proposed. The fluid inside the colon is first identified
and removed based on its characteristic of horizontal surface.
A simple 3D region growing algorithm is applied to obtain
initial segmentation of air, which is served as the basis of the
ensuing automatic locating object and background seeds. Then
the isolated-connected threshold algorithm, together with the
above seeds, is applied to obtain the final results. The colon can
be obtained by applying morphological operations to the
segmentation results of air. Our proposed algorithm can
automatically segment different substances based on the
isolated-connected threshold. It allows the user to modify the
segmentation results interactively by providing more object or
background seeds.
Keywords – virtual colonoscopy, colon segmentatioin,
isolated-connected thresholding, image processing, region
growing
I.
INTRODUCTION
Colorectal cancer has been the third leading cause of
cancer deaths in Taiwan for the last decade[1]. Even though
the exact cause of most colorectal cancers is still not known,
it is possible to prevent many cases. Prevention and early
detection are possible because most colorectal cancers
develop from polyps – precancerous tissue growths. Early
screening can help find polyps, which can be easily removed,
thereby lowering a person’s cancer risk.
For years, optical colonoscopy (OC) has been the gold
standard for colorectal cancer screening. However, OC is
often regarded as an invasive, highly uncomfortable and
expensive technique, and thus becomes an undesirable
procedure patients are reluctant to undergo. An alternative
diagnostic procedure to OC is virtual colonoscopy(VC),
which is the process of combining multi-slice CT images and
advanced visualization techniques to allow radiologists to
interactively view, manipulate and examine the interior of
the colon and even detect polyps or tumors. VC is a quick
and non-invasive procedure that does not require sedation or
anesthesia. It eliminates physical discomfort and associated
risks, such as bleeding and perforation of the colon wall.
Colon segmentation is an essential step of creating a 3D
colon model for VC. To become a clinically viable screening
method, VC needs automatic segmentation methods to
reduce the amount of user interaction required to generate
accurate models. Several sophisticated semi- and fully
automated segmentation algorithms for the colon have been
reported in the literature, which are commonly based on
region- or volume-growing techniques. Wyatt et al. [2] used
a region-growing technique to segment the colon, in which
the automated selection of seed voxels was based on the
distance transform. Although the colon was segmented
satisfactorily, a small bowel or stomach was present in a
majority of the segmentation results. Chen et al. [3] used
vector quantization to label voxels. The colonic walls were
segmented by region growing based on the labeled voxels. 6
of 21 datasets cannot be segmented satisfactorily. Masutani
et al. [4] used an anatomy-oriented technique to remove
anatomic structures surrounding the colon before
thresholding of the colon wall. This approach was later
combined with a region-growing-based technique for final
segmentation. However, 10%-15% of the results consisted of
extracolonic objects, such as small bowel and stomach.
Three major problems makes colon segmentation a
difficult task to automate [2][5]. First, the colon in a CT
image often consists of an uncertain number of disconnected
regions due to its complex 3D winding structure. Second, the
colon is not the only air-filled structure in the abdomen.
Lower portions of the lung are often present, and portions of
the small bowel and stomach may also be partially filled with
air. It is incorrect to claim that all air voxels are part of the
colon. Third, obstructions in the colon itself complicate
automated segmentation, prevent a continuous colon
segment, and require the use of multiple manually placed
seed points. Possible obstructions include fluid, very large
lesions, and residual feces. The existence of such
obstructions creates the partial-volume-effect(PVE)[6],
which emerges on the boundary between low and high
density regions during imaging.
In clinical application, the physician may expect to have
an opportunity to modify the results for verification purposes,
both interactively and intuitively. In this paper, an interactive
colon segmentation algorithm is proposed. The algorithm is
essentially an isolated-connected threshold operation with
seed points determined automatically, not interactively. The
physician is then allowed to examine and modify the result
intuitively - simply click the wanted or unwanted regions.
The remainder of this paper is organized as follows.
Section 2 describes the segmentation procedure using
isolated-connected threshold. And several experimental
results are given in Section 3, followed by a conclusion in
Section 4.
II.
Fig. 1. (a) a clipped colon image (b) air component
(c) fluid component
METHODOLOGY
The proposed colon segmentation method is based on
identifying the colonic interior, i.e. the air inside the colon,
first. And based on this interior identification, the colon can
consequently be segmented out. The overall segmentation
procedure consists of the following steps: pre-processing,
preliminary segmentation, isolated-connected threshold and
colon identification.
A. Pre-processing
In clinical practice, it is quite difficult to remove the fluid
completely during bowel preparation. The fluid can separate
the colon passage into several disconnected volumes,
especially when the segmentation proceeds based on the air
channel. Thus, the pre-processing procedure aims at
removing fluid inside the colon digitally.
1. Air and Fluid Identification
A proper threshold value is selected based on the graylevel histogram of the whole volumetric abdominal CT
dataset to differentiate the air and fluid inside the colon. An
example of the result is illustrated in Fig. 1. The identified
fluid components will be removed from the image – labeling
fluid voxels as air.
2. PVE Voxels Removal
In this research, we focus on the PVE on the air and fluid
boundary, where the intensities of these voxels do not belong
to either the air or the fluid ranges. When CT images are
acquired, the fluid inside the colon gathers into some
concave parts of the colon and has a horizontal surface due
to the gravity force.
We use the vertical filter technique to remove PVE
voxels[6]. The colon passage become connected, as
illustrated in Fig. 2, and the subsequent 3D region growing is
therefore viable.
Fig. 2. Segmentation results (a) before (b) after
removing PVE voxels
Fig. 3. Automated seed point selection
based on anatomical knowledge
Fig. 4 Utilize the bounding box to determine
the first set of seed points.
B. Preliminary Segmentation
The preliminary segmentation is achieved by applying
the 3D region growing. This preliminary result is used to
help determine seed sets automatically required in Step C.
1. Seed Point Placement
The seed point for 3D region growing can be localized
without human involvement, based on the fact that the last
slice of the whole volumetric data contains the rectum.
Otsu’s threshold scheme and a logical operation between
images are performed to extract the hollow area of the
Fig. 5. Search for the second set of seed points. The search
area is limited to the dash box for the left-top quadrant.
tube[7][8]. The center pixel of the segmented region is used
as a seed voxel for 3D region growing, as shown in Fig. 3.
2. 3D Region Growing
Since the initial seed point is automatically located and
the PVE voxels are removed, a 3D region growing algorithm
can be applied to the whole colon dataset for air
segmentation. This well-known algorithm is described in
many texts and has several variations [9].
We use a similarity measure to determine if voxels
belong to the region. Given an initial seed point, the
neighborhood is examined to determine if any connected
voxels have similar characteristics to the seed point. Those
voxels in the neighborhood with a sufficient similarity
measure are collected. Since the air component has a certain
gray-level value, i.e. -1000 HU, a more strict similarity
measure is preferred. This process continues by removing a
voxel from the collected neighbors, adding it to the region,
and examining its neighbors for placement in the collection.
The algorithm stops when no more voxels remain in the
collected neighbor set.
for radiologists to explore not only the colon surface, but
also its underneath tissues
III.
EXPERIMENTAL RESULTS
The proposed algorithm was implemented based on the
Insight Segmentation and Registration Toolkit (ITK), which
provides extensive C++ classes for image analysis[10]. We
applied our method to several cases of virtual colonoscopy.
Examples of CT images and their segmentation results are
illustrated in Fig. 6.
In Fig. 6, images in column (a) are clips from the original
CT images for better visualization effects. Their preliminary
results, i.e. after applying a 3D region growing algorithm
with a gray-level range of -1024 to -475, are shown in
column (b). It is easy to see that the results are unsatisfactory.
Based on results in column (b), we apply isolated-connected
threshold algorithm to further improve the results, as
3. Seed Sets Placement
Based on the result of 3D region growing, two sets of
seed points are collected automatically for later use.
For each region in a 2D image, all pixels lies on its
bounding box are labeled as the first set of seeds, as shown
in Fig. 4.
For the second set of seeds, a search process is conducted
for each quadrant of a region, which starts from the center of
the bounding box. The search area for each quadrant is as big
as the bounding box, as illustrated in Fig. 5. During the
search process, once a pixel, whose value falls between the
air and colon, is found, the pixel is added to the second set of
seeds, and the search stops. Therefore, for each region, there
are four pixels selected as the seeds, one for each quadrant.
C. Isolated-Connected Threshold
The isolated-connected threshold uses a binary search to
adjust lower and upper values, trying to find the optimal
threshold to ensure that all of the first seeds are contained in
the resulting segmentation and all of the second seeds are not
contained in the segmentation. [10].
Since two sets of seed points are obtained in Step B-3,
the isolated-connected threshold algorithm can be executed
without user involvement. The user can examine the
segmentation results with little effort. And if necessary, the
user can modify segmentation results in an intuitive way simply using mouse clicking to label pixels as ‘wanted’ or
‘unwanted’ to change the two sets of seed points. The
isolated-connected threshold is then performed again to
reflect such a change of seed points.
D. Colon Segmentation
The segmentation results from the previous step actually
are composed of air and removed fluid components, not the
colon itself. A 3D binary morphological operation - dilation
is applied to such results to obtain the colon itself [11]. The
greater size of the dilation operator is, the more tissue voxels
contained beneath the colon surface, which makes it possible
Fig. 6. (a) original images; (b) preliminary segmentation
results; (c) final segmentation results.
Air/Fluid components are put in red color.
illustrated in column (c). It is approved by the radiologist
that images in column (c) have better segmentation results.
IV.
CONCLUSION
In this paper, we have proposed an automatic colon
segmentation algorithm. After removing fluid inside the
colon digitally in the pre-process stage, the threshold-based
3D region growing is applied to obtain preliminary results.
And the isolated-connected threshold is also applied to get
better results thereafter. The selection of seed points for both
3D region-growing and isolated-connected threshold can be
carried out automatically. Thus, the radiologist can easily
view the segmentation results with little effort.
Besides the automatic processing flow, our method also
allows the radiologist to refine the results in an intuitive
approach at the final stage. This interaction is quite helpful in
clinical use.
Our future works include a quantitative evaluation, and a
clinical validation of the proposed algorithm.
ACKNOWLEDGMENT
This work was supported by the grant contract NSC 992221-E-212-012-MY3 from National Science Council,
Taiwan, R.O.C.
REFERENCES
[1]
Department of Health. "Leading Causes of Death in Taiwan in 2010."
Health 99 Education Resource. http://health99.doh.gov.tw/
Hot_News/h_NewsDetailN.aspx?TopIcNo=6240 (16 July, 2011)
[2] C.L. Wyatt, Y. Ge, D.J. Vining, "Automatic Segmentation of the
Colon for Virtual Colonoscopy," Computerized Medical Imaging and
Graphics, Vol. 24, No. 1, 2000, pp. 1-9.
[3] Dongqing Chen, Zhengrong Liang, Mark R. Wax, Lihong Li, Bin Li,
and Arie E. Kaufman, "A Novel Approach to Extract Colon Lumen
from CT Images for Virtual Colonoscopy," IEEE Transactions on
Medical Imaging, Vol. 19, No. 12, December 2000, pp. 1220-1226.
[4] Yoshitaka Masutani, Hiroyuki Yoshida, Peter M. MacEneaney,
Abraham H. Dachman, "Automated segmentation of colonic walls for
computerized detection of polyps in CT colonography," Journal of
Computer Assisted Tomography, Vol. 25, No. 4, July/August 2001,
pp. 629-638.
[5] Dongqing Chen, Rachid Fahmi, Aly A. Farag, Robert L. Falk, and
Gerald W. Dryden, "Accurate and fast 3D colon segmentation in CT
colonography," IEEE International Symposium on Biomedical
Imaging: From Nano to Macro, Boston, Massachusetts, U.S.A., June
28-July 1, 2009, pp.490 - 493.
[6] M. Sato, S. Lakare, M. Wan, A. Kaufman, Z. Liang, and M. Wax,
"An automatic colon segmentation for 3D virtual colonoscopy,"
IEICE Transactions on Information and Systems, Vol. E84-D, No. 1,
January 2001, pp. 201-208.
[7] Nobuyuki Otsu, "A Threshold Selection Method from Gray-Level
Histograms," IEEE Transactions on Systems, Man and Cybernetics,
Vol. 9, No. 1, Jan. 1979, pp. 62-66.
[8] Chien-Chuan Ko and Jyh-Wei Jang, "Interactive Polyp Biopsy Based
on Automatic Segmentation of Virtual Colonoscopy," IEEE
Symposium on Bioinformatics and Bioengineering, Taichung,
Taiwan, 19-21 March, 2004, pp. 159-166.
[9] Rafael C. Gonzalez and Richard E. Woods, Digital Image Processing,
Reading, MA:Addison-Wesley, 3rd edition, 2008.
[10] Luis Ibanez, Will Schroeder, Lydia Ng and Josh Cates, "The ITK
Software Guide: The Insight Segmentation and Registration Toolkit,"
U.S.A., Kitware Inc., 2005.
[11] Wook-Joong Kim, Seong-Dae Kim, and Hyder Radha, "3D Binary
Morphological Operations Using Run-length Representation," Signal
Processing: Image Communication, Vol. 23, No. 6, July 2008, pp.
442-450.