Liver Segmentation from CT Images Based on

Region Growing Method

Yufei Chen1,2,3, Zhicheng Wang1,2, Weidong Zhao1,2, Xiaochun Yang1,2
1
Research Center of CAD, Tongji University, Shanghai 200092, China
2
The Engineering Research Center for Enterprise Digital Technology, Ministry of Education, Tongji University, Shanghai
200092, China
3
Fraunhofer Institute for Computer Graphics, Cognitive Computing & Medical Imaging, 64283 Darmstadt, Germany
[email protected]

Abstract—Accurate liver segmentation on computed tomography are the most common used priori knowledge [5, 6]. They
(CT) images is a challenging task because of inter and intra- always combine with other approaches. Statistical based
patient variations in liver shapes, similar intensity with its nearby approaches: A statistical model discrimination of the liver is
organs. We proposed a liver segmentation method based on established from quantities of data sets, and then the model is
region growing approach. First of all, basic theory of region used to pre-process the images and obtained liver likelihood
growing approach is introduced. Secondly, a pre-processing images for further process [7, 8, 9]. Nevertheless, the model
method using anisotropic filter and Gaussian function is generation is time consuming and could not respond sensitively
employed to form liver likelihood images for the following among different patients. Active contour models: It is the most
segmentation. Thirdly, an improved slice-to-slice region growing
popular used method in liver segmentation, including fast
method combined with centroid detection and intensity
marching [10], level set method [11, 12], snake model [13, 14,
distribution analysis is proposed. Finally, the superior liver
region is extracted by applying the morphologic operation. 15, 16]. But they are useful only when a good initial estimate is
Experiments on a variety of CT images show the effectiveness present. Most of the models compound the computing time and
and efficiency of the proposed method. get unsatisfied results for the slices with fuzzy liver boundary.
We present a segmentation method based on improved
Keywords-liver segmentation; anisotropic filter; Gaussian region growing. The paper is organized as follows. In section
function; region growing method; centroid detection; morphologic
II, the region growing method is introduced, the principle
operation
methods and theories in our liver segmentation process are
described in section III. The experimental results are presented
I. INTRODUCTION in section IV. Finally the paper concludes in section V.
Abdominal CT images have been widely studied in recent
years. Some of the current interests are the automatic diagnosis II. REGION GROWING METHOD
of liver pathologies and three-dimension rendering. The first
Region growing method has been widely used in image
and fundamental step in all these studies is the automatic liver
processing [17, 18, 19], which segments images into several
segmentation, which is still a complex problem [1].
homogeneous regions based on a seed point set [20].
There are several reasons make the accurate segmentation
in liver difficult. Firstly, the image is always very noisy and the A. Basic Theory
partial volume effect causes the liver boundary ambiguous. As the name implies, region growing process begins with
Secondly, the overlaps result in gaps and cavities. Thirdly, the seed point set and grows by adding neighbor pixels that
there are large variations in human liver geometric properties satisfy the similarity constraint. This process is repeated until
like liver size and shape between patients, this can be worse if all pixels belong to some region, as illustrated in Fig. 1.
the patient has liver operations before due to alcoholic cirrhosis
or liver tumor. Besides, even the same organ in the same
patient may exhibit different intensity values. Finally, the
nearby organs such as right kidney, stomach, spleen and
abdominal wall, have similar intensity values with liver, which
makes it harder to extract liver only. A few literatures and
algorithms about this research topic with high automation have
been proposed. They can be categorized into four groups:
Intensity based approaches: The most common procedure is
to apply threshold operators to discard regions with intensity
outside the liver range [2, 3, 4]. But the thresholds affect the
Figure 1. Region growing process.
result directly and hard to determine. Priori knowledge based
approaches: The topological, distance and orientation relations

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 13,2023 at 13:01:23 UTC from IEEE Xplore. Restrictions apply.
978-1-4244-2902-8/09/$25.00 ©2009 IEEE 1
 B. Consideration Issues −
| I 0 − μ |2
I = I0 ⋅e 2σ 2 (2)
Region growing method often gives out good segmentation
result related with two important issues:
1) How to choose the seed point set? I greatly reserves the primal intensity information of liver
and reduces the effect of non-liver regions. The determination
2) How to choose the similarity constraint? method in μ and σ is discussed in the following sections.

III. DESCRIPTION OF OUR METHOD B. Region-Growing

Generally speaking, the measure standards of liver In order to get a faster and more accurate result, the liver
segmentation performance mainly focus on method robustness, segmentation method always combines with priori knowledge
result accuracy and processing speed. that focuses on the liver location, size, intensity distribution
and topological relations between slices. We make full use of
A. Image Pre-processing the topological continuity between adjacent CT slices in one
data set to form the criterion in slice-to-slice region growing.
Anisotropic filter is a diffusion process, which can be used
The continuity of liver region in centroid coordinate and
as a preprocessing step to enhance morphological definition of
intensity distribution is taken into account: centroid coordinate
the input image by sharpening discontinuities, and remove
of liver region has a little change between slices that can be
noise in homogeneous regions while preserving object
used as site guide in the following slice; the mean and standard
boundaries and fine details [21].
deviation values of liver region intensity from the previous
slice give a better estimation of liver intensity distribution on
⎧ current slice.
⎪ I t ( x, y, t ) = ∇ ⋅ C (| ∇I ( x, y, t ) |)∇I ( x, y, t )
⎪ For each CT data set, assume it has q slices together, in the
⎪⎪ −
|∇I |2 nth slice, the region growing process starts from a seed point
⎨C (| ∇I |) = e 2 K
2
(1) SPn, which is auto obtained from the result of previous slice.
⎪ Centroid detection technology is applied to get its coordinate
m m
⎪ 1 noted as (xcn-1, ycn-1). Only the seed point on the first slice (x0,
⎪
⎪⎩
K = C ⋅ 2 ∑∑
m y =1 x =1
| ∇I ( x , y , t ) 2 | y0) should be predetermined artificially.

where I(x, y, 0) is the input image, t refers to the iteration steps. ⎧⎪SP n = ( x cn −1 , y cn −1 ) n = 2, " q
The diffusion function C depends on the gradient magnitude, ⎨ 1 (3)
while parameter K controls the diffusion process and changes ⎪⎩SP = ( x 0 , y 0 ) n =1
with the gradient magnitude adaptively, where m =3 in our
experiment represents the neighbor’s size. Let An stands for the liver result region A of slice n, the
The Gaussian function is introduced to make the output general function for calculating the centroid (xc, yc) is:
image after the filter, noted as I0, more adaptive for liver
segmentation. In mathematics, the graph of a Gaussian is a
⎧
characteristic symmetric bell shape curve that quickly falls off
towards minus infinity. As Fig.2 shown, the parameter μ is the ⎪⎪ n ∫
xc = A n
n
x d An
, y cn =
∫ An
y d An
n = 2, " q (4)
position of the center of the peak, and σ controls the width of ⎨ A An
⎪ 1
the bell. y 1c = y 0
⎩⎪ x c = x 0 , n =1

Then, region is grown from the seed point SPn by adding

neighbor pixels whose intensity In(x, y) satisfy the similarity
constraint given in Equation (5):

| I n ( x, y ) − μ sn | < σ sn s = 1,2, " p, n = 1,2, " q (5)

where In(x, y) means the intensity value of pixel (x, y) on the

Figure 2. Gaussian function.
nth slice. s stands for some growing step in each slice, with p
as the total number. μ sn represents the average intensity value
Here, we use it to describe the membership degree. For of the grown region in the sth growing step on the nth slice,
pixels with intensities in range of [μ -σ, μ +σ], they have the while σ sn is the corresponding standard deviation value. The
higher probabilities as inside the liver part. Otherwise, they are
deemed as the non-liver pixels. Therefore, the liver-likelihood computing formulas can be written as:
image I can be computed as:

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 13,2023 at 13:01:23 UTC from IEEE Xplore. Restrictions apply.
2
 N N IV. EXPERIMENTAL RESULTS
1 1
μ ns =
N ∑I n
( x i , y i ), σ sn =
N ∑ (I n
( x i , y i ) − μ sn ) 2 (6) We performed our experiment on a variety of CT images.
i =1 i =1 The results showed that the treated images had fast and reliable
segmentations by our method, as shown in Fig. 4. (a) (b) (c) (d)
where N is the pixel number located inside the grown region. are the segment result with the slice number respectively as:
μ1n and σ 1n for the first region growing step are auto n=10, n=11, n=20, n=21 in one data set. The left ones show the
region growing results on the original images, with the results
obtained from the previous processed slice. As in Equation (7), extracted on the right. (e) illustrates the liver surface rendering
they are predetermined artificially only in the first slice, where on the segmentation result. The proposed method also
I1(x0, y0) is the intensity value of pixel (x0, y0): performed better when it came to efficiency: The computing
time T of each slice from (a) to (d) were 3.553s, 3.244s, 2.164s,
⎧⎪μ1n = μ np −1 , σ 1n = σ np −1 and 2.113s.
⎨ 1 (7)
⎪⎩μ1 = I 1 ( x0 , y0 ), σ 11 = 20 Our experiment was performed on CT data sets stored as
DICOM images, 512 by 512 pixels with a 12-bit gray level
resolution and 5mm interval. They were offered by the East
The growing process is repeated until no other pixels can Hepatobiliary Surgery Hospital in Shanghai, China. The
be added. computer used for runtime measures had a Pentium (R) D CPU
(2.8GHz), 1GB of RAM and a Windows XP operating system.
C. Image-Postprocessing
The region growing result may contain holes and connected
neighbor tissues. Mathematic morphologic operation, the
opening operation, is performed with a structure element of
radius=3. The value can be artificially changed for specific
application. Then, the final grown region A, i.e. liver result, is
obtainable by removing the fragments and reserving the main
part, then filling the holes of it. Calculate μ np and σ np according
to Equation (6) for the follow-up slice use in (2) and (7).

D. Liver Segmentation Steps

To summarize, the principal steps of the method are (a) n=10 T=3.553s
illustrated in Fig. 3:

Image Reading n=1

Seed Point (x0, y0)
Anisotropic Filter

Image-Preprocessing Gaussian Functionn

Seed Point
SPn
Region-Growing

Similarity
(b) n=11 T=3.244s
Constraint

Image-Postprocessing Morphologic Operation

, , SP
Parameter Computing

N Y
The End of Slice? Surface Rendering

(c) n=20 T=2.164s

Figure 3. Diagram of the proposed method.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 13,2023 at 13:01:23 UTC from IEEE Xplore. Restrictions apply.
3
 [3] Y. Zheng, X. Yang, X. Ye, X. Lin, “Fully Automatic Segmentation of
Liver from Multiphase Liver CT,” Medical Imaging 2007: Image
Processing, SPIE, vol.6512, pp. 6512X, 2007.
[4] T. Saitoh, Y. Tamura, T. Kaneko, “Automatic Segmentation of Liver
Region Based on Extracted Blood Vessels,” Systems and Computers in
Japan, vol.35, pp.1-10, 2004.
[5] P. Campadelli, E. Casiraghi, G. Lombardi, “Automatic Liver
Segmentation from Abdominal CT Scans, ” 14th International
Conference on Image Analysis and Processing-ICIAP 2007, IEEE
Computer Society, pp. 731-736, 2007.
[6] J. B. Fasquel, V. Agnus, J. Moreau, L. Soler and J. Marescaux, “An
Interactive Medical Image Segmentation System Based on the Optimal
Management of Regions of Interest Using Topological Medical
Knowledge, ” Computer Methods and Programs in Biomedicine, vol.82,
(d) n=21 T=2.113s pp.216-230, 2006.
[7] X. Zhou, T. Kitagawa, K. Okuo, T. Hara, H. Fujita, “Construction of a
Probabilistic Atlas for Automated Liver Segmentation in Non-contrast
Torso CT Images”, International Congress Series, vol.1281, pp.1169-
1174, 2005.
[8] T. Okada, R. Shimada, Y. Sato, M. Hori, K. Yokota, M. Nakamoto, Y.
W. Chen, H. Nakamura, “Automated Segmentation of the Liver from 3D
CT Images Using Probabilistic Atlas and Multi-level Statistical Shape
Model,” MICCAI 2007, LNCS, vol.4791, pp. 86-93, 2007.
[9] H. Lamecker, T. Lange, M. Seebaβ, S. Eulenstein, M. Westerhoff, H. C.
Hege, “Automatic Segmentation of the Liver for Preoperative Planning
of Resections,” Stud Health Technol Inform, vol.94, pp.171-173, 2003.
[10] J. Varma, J. Durgan, K. Subramanyan, “A Semi-automatic Procedure to
Extract Couinaud Liver Segments from Multi-slice CT Data,”
Visualization, Image-Guided Procedures, and Display, SPIE 2003,
(e) vol.5029, pp. 561-568, 2003.
[11] S. Pan, B. M. Dawant, “Automatic 3D Segmentation of the Liver from
Figure 4. Some liver segmentation results. Abdominal CT Images: a Level-set Approach,” Medical Imaging 2001:
Image Processing, SPIE 2001, vol.4322, pp. 128-138, 2001.
Future work consists in liver vessels segmentation and [12] J. Lee, N. Kim, H. Lee, J. B. Seo, H. J. Won, Y.M. S, Y. G. Shin, S. H.
representation, volume rending realization of liver including Kim, “Efficient Liver Segmentation Using a Level-set Method with
vessels and tumors, volume measurement of liver region, visual Optimal Detection of the Initial Liver Boundary from Level-set Speed
surgery of liver disease. Images,” Computer Methods and Programs in Biomedicine, vol.88,
pp.26-38, 2007.
[13] A. Evans, T. Lambrou, A. Linney, A. Todd-Pokropek, “Automatic
V. CONCLUSIONS Segmentation of Liver Using a Topology Adaptive Snake,” the IASTED
International Conference on Biomedical Engineering, Acta Press, pp.
In this paper, we propose an effective method for automatic 205-208, 2004.
liver segmentation from CT images based on region growing
[14] J. Tian, F. Zhu, X. Luo, “Medical Image Segmentation Using Improved
approach. The method takes the image sequence continuity on Active Contour Model,” Third International Conference on Photonics
topology and intensity into account, combines region growing and Imaging in Biology and Medicine, SPIE 2003, vol.5254, pp. 176-
method with centroid detection and intensity analysis. Besides, 185, 2003.
morphologic operation is applied to guarantee a fast and [15] S. Huang, B. Wang, X. Huang, “Using GVF Snake to Segment Liver
accurate liver region. The experimental results of the liver from CT Images,” International Summer School and Symposium on
segmentation demonstrate the efficiency and effectiveness of Medical Devices and Biosensors, MIT, Boston, IEEE-EMBS, pp. 145-
148, 2006.
the proposed method. We can conclude that our method gives
[16] Y. Seol, J. Yu, T. Kang, “Resolving the Initial Contour Problems of
out efficient and effective performance for liver segmentation GVF Snake in the Sequential Images,” 29th Annual International
on CT images. Conference, Lyon, France, IEEE EMBS, pp. 779-782, 2007.
[17] M. MANCAS, B. GOSSELIN, B. MACQ, “Segmentation Using a
ACKNOWLEDGMENTS Region Growing Thresholding,” Image Processing: Algorithms and
Systems IV, Proceedings of the SPIE, vol.5672, pp.388-398, 2005.
This work was supported by program for Young Excellent [18] Z. Lin, J. Jin, H. Talbot, “Unseeded Region Growing for 3D Image
Talents in Tongji University (No.: 0800219049). Segmentation,” Pan-Sydney Workshop on Visualisation, vol.2, pp.31-
37, 2000.
[19] Y. Chen, C. Han, “A Modified Region Growing Algorithm for Multi-
REFERENCES colored Image Object Segmentation,” Chinese Optics Letters, vol.5,
[1] E. Casiraghi, G. Lombardi, S. Pratissoli, S. Rizzi, “3D α-Expansion and pp.25-27, 2007.
Graph Cut Algorithms for Automatic Liver Segmentation from CT [20] R. Adams, L. Bischof, “Seeded Region Growing,” IEEE Trans.Pattern
Images,” KES 2007, Part I, LNAI, vol.4692, pp.421-428, 2007. Anal.Mach.Intelligence, vol.16, pp. 641-647, 1994.
[2] L. Gao, D. G. Heath, B. S. Kuszyk, E. K. Fishman, “Automatic Liver [21] P. Perona, J. Malik, “Scale-space And Edge Detection using Anisotropic
Segmentation Technique for Three-dimensional Visualization of CT Diffusion,” IEEE Trans. Pattern Anal.Mach.Intelligence, vo1.2, pp. 629-
Data,” Radiology, vol.201, pp. 359-364, 1996. 639, 1990.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 13,2023 at 13:01:23 UTC from IEEE Xplore. Restrictions apply.
4