Вы находитесь на странице: 1из 6

2015 International Conference on Industrial Instrumentation and Control (ICIC)

Col/ege ofEngineering Pune, India. May28-30,2015

Analysis of the liver in CT images using an improved


region growing technique

P.Arjun M.K.Monisha
Dept. of Electronics Engg., Dept. of Electronics Engg.,
MIT Campus, Anna University, MIT Campus, Anna University,
Chennai, India Chennai, India
arj un. parthipan93@gmai1.com monishamk94@gmai1.com

A.Mullaiyarasi G.Kavitha
Dept. of Electronics Engg., Dept. of Electronics Engg.,
MIT Campus, Anna University, MIT Campus, Anna University,
Chennai, India Chennai, India
mullai.ilasu@gmail.com kavithag_mit@annauniv.edu

Abstract-This paper presents an improved region growing related structures for various abnormalities when other types
algorithm to enhance the segmentation of the liver from of examinations are not conclusive. It is also used for cancer
abdominal CT images. The abdominal CT images are detection and to provide guidance for biopsies. Image
characterized by poor contrast and blurred edges which increase
segmentation plays an important role in medical image
the complexity of liver segmentation. Initially, the images are
analysis. An accurate segmentation of the liver is crucial in
subjected to preprocessing which involves de-noising,
thresholding and non-linear mapping. Then, the improved region
clinical diagnosis and study. Pre-processing of the liver CT
growing algorithm is applied to the preprocessed liver images. images is required for accurate liver segmentation. The
Post processing is performed using a combination of extraction of liver anatomical information from the abdominal
morphological operations. The results of the improved algorithm CT image requires a lot of expertise and is time consuming.
are compared with the traditional region growing algorithm and Several algorithms have already been implemented to
the k-means clustering algorithm to show the effectiveness of the segment the liver from the abdominal CT image. Shraddha
proposed method. Performance validation is also done by Sangewar et a1. [2] performed liver segmentation using the
comparing the results with the ground truth. Similarity measures
modified K-means method along with specialized contouring
namely the Dice similarity, Sokal and Sneath-I similarity, Sokal
algorithm. Suhuai Luo et al. [3] proposed liver segmentation
and Sneath-II similarity and Tanimoto similarity are used for the
comparison. The results obtained using the improved method
using wavelets and machine learning. Hans Bhurkhardt et al.
give an accuracy of 97%. The average Dice similarity measure [4] applied a hybrid segmentation algorithm to perform liver
for the considered images was found to be 0.86. The average segmentation. Other methods include the Adaptive fast
correlation coefficient between the ground truth and the marching method [5], local entropy based method [6], region
segmented result are also high in the improved algorithm. The scalable fitting model [7], pseudocolorization approach [8],
obtained results seem to be clinically relevant. statistical model [9] and graph partitioning approaches [10].
However, great challenges still remain in liver extraction on
the aspects of accuracy, automation and robustness.
Keywords- Liver segmentation, non linear mapping, improved
In this work, an improved region growing algorithm is
region growing, k-means clustering
implemented to enhance the accuracy of liver segmentation.
The pre processing is performed using a Gaussian high boost
I. INTRODUCTION
filter followed by threshold segmentation and non linear
The liver is the largest internal organ in the human body. It mapping. Then the improved region growing algorithm is
performs a diversity of life sustaining functions and virtually applied, followed by morphological operations in the post
affects every physiological process in the body. It acts as a processing stage. The obtained results are compared with the
blood purifier, nutrient processing unit and also controls the traditional region growing algorithm and the k-means
metabolism in the body. It primarily controls the red blood cell clustering algorithm. Performance evaluation is carried out
decomposition, plasma protein synthesis and hormone using statistical measures such as sensitivity, specificity and
production. Computerized tomography processes a series of accuracy. The correlation coefficient between the ground truth
X-ray views taken from different angles to create cross and the segmented area is also calculated for all three
sectional images of the bones and soft tissues inside the body. methods. Further, similarity measures are used to validate with
The CT scan of the liver is used to assess the liver and its the ground truth.

978-1-4799-7165-7/15/$31.00 ©2015 IEEE 1561


II. PROBLEM SPECIFICATION segmentation, it is a challenging task to differentiate the liver
The liver needs to be segmented from the Abdominal CT from the surrounding organs with similar gray levels.
image. The difficulty in segmentation is in differentiating the Grayscale transformation enhances the contrast of the
liver from the surrounding organs in the abdomen. The CT abdominal CT image to obtain the features in the image.
images are characterized by low contrast, the liver shares Hence the non-linear mapping technique is introduced to
similar gray levels to that of its surrounding organs. The enhance the image contrast in this work. The sigmoid filter is
complexity and wide variety of liver shapes further increases used for the transformation. It is given by:
the difficulty of the liver segmentation task. Also, 1

abnormalities in the liver needs to be detected based on the p' = (Max - Min) �+ M'III
(3)
l+e a
minor variations in the gray level values.

where P is the gray level of input pixel, P' is the gray level of
III. METHODOLOGY output pixel, Min and Max are the minimum and maximum
gray level of the output image, a defines the width of the
input gray level range and � defines the gray level around
which the range is centered. This method enhances the
contrast of input CT image.

B. K-means clustering algorithm


The K-means algorithm [2] classifies a data set through a
certain number of clusters. It involves defining k centroids for
k number of clusters and associating each point in the dataset
to the closest centroid. A loop is generated and the k centroids
change their location step by step until no more changes are
possible. The objective function called the squared error
function needs to be minimized. The squared error function is
given by:
Figure l. Flow chart of the proposed work �k �n
Xi1
-L... i =lL...i=l(
J -
_
C1)2 (4)

A. Pre-processing
where
( l
X i
_
C1)2 is a chosen distance measure between a
Figure 1 shows the flow of the proposed work. The
preprocessing stage involves denoising, thresholding and non data point xl and the cluster centre c1 and is an indicator of
linear mapping. The CT images are characterized by the the distance of the n data points from their respective cluster
presence of Gaussian noise acquired during the data centers.
acquisition process. Hence image de-noising is performed by
using a Gaussian high boost filter. The high boost filter
C. Improved region growing algorithm
emphasizes the high frequency components without
eliminating the low frequency components. Thus the Gaussian After pre-processing, the improved region growing
high boost filter is used for eliminating the noise effects as algorithm is used for liver segmentation. It is classified as a
well as sharpening the image. pixel-based image segmentation method since it involves the
selection of initial seed points. A quasi random sequence of
The Gaussian high pass filter is given by seed points is generated and the region growing criteria is thus
improved. The basic function of this technique is to partition
(1)
an image into non overlapped regions. It takes seed points as
input, and then merges the pixels with similar property and
where D(u,v) .Ju2 +
= v2 is the distance from the origin in produces a region corresponding to the seed point.
the frequency plane.

The output g(m,n) of a high boost filter is

gem, n) = (A - l)f(m, n) + highpass(f(m, n)) (2)

where f(m,n) is the original image and A is a constant.


The second step in pre-processing involves threshold
segmentation in order to remove the high gray level organs
such as the spinal cord and the ribs which reduces the Figure 2. The region of interest and selection of seed points in the liver
region
complexity of the segmentation process. After threshold

1562
Initially, a region of interest(ROI) is set in the liver region The formulae for the performance measures are as follows:
of the preprocessed image using priori knowledge. Then,
TP+TN
multiple seed points are generated in the region of interest Accuracy = (9)
TP+TN+FP+FN
using the Halton' s quasi random sequence. The seed points
generated in the ROI is shown in figure 2. The gray level of
TP
each of these seed points is measured. If the gray level is 0 Sensitivity (10)
TP+FN
caused due to holes after pre-processing, these points are
called singular points and are eliminated. The remaining seed
TN
points are used for calculating the grayscale average in the Specificity (11)
TN+FP
ROI. The grayscale average is given by

N�nIf=-;.n Gi
Ga 2TP
=
(5) Dice similarity coefficient = (12)
2TP+FP+FN

(6) Sokal and Sneath-I TP


similarity coefficient = TP+2FP+2FN (13)
Where Ga is the grayscale average in the ROI, Gi is the gray
level of the pixel, N is the total number of seed points Sokal and Sneath-II 2(TP+TN)
generated, n is the number of singular points. A large number similarity coefficient = 2TP+FP+FN+2TN (14)
of seed points may be generated but the number of seed points
is reduced to lower the amount of computation. The points Tanimoto TP
which meet the formula (6) are chosen as seed points where x similarity coefficient = (TP+FP)+(TP+FN)-TP (15)
is a percentage whose value depends on the number of seed
points. where TP is the true positive value, TN is the true negative
The region growing criteria is set using the gradient of value, FP is the false positive value and FN is the false
the pixel. The Robert operator is used for calculating the negative value.
gradient of pixels corresponding to the seed points. The
average gradient in the ROI is computed by the formula:
IV. RESULTS AND ANALYSIS
N-n
VF =
N-l-
-nI _ VF.
[-1 [
(7) The CT images considered for the analysis were obtained
from the Cancer Imaging Archive. 40 normal and abnormal
where VF is the average gradient and VF;. is the gradient of images were considered for the analysis of the proposed
the pixel. algorithm. Figure 3(a) shows the raw abdominal CT image. It
was found that the abdominal CT images were composed of
The region growing criteria is given by: similar gray levels throughout. Hence the pre-processing
stages were introduced. Figure 3 shows the various steps
(8) involved in the preprocessing stage. The de-noising and
threshold segmentation stage sharpened the image and
where a and b are control parameters. The gray level of the removed the high gray level organs from the abdominal CT
pixel in the growing region lies in a fixed range which is respectively. Figure 3(b) shows the image after denoising
dependent on the grayscale average and gradient average. using the Gaussian high boost filter. Figure 3(c) shows the
image after thresholding to remove the high gray level organs.
D. Post-processing
The non linear mapping using the Sigmoid function resulted in
The segmented liver result may contain holes or be over the distribution of gray levels over a wide range. Figure 3(d)
segmented. Hence a post processing step is required to obtain shows the image after applying non linear mapping. The liver
a smooth liver contour devoid of holes. The majority filling region is easily distinguished and the improved region
algorithm is implemented. According to this algorithm, a pixel growing algorithm is applied.
is set to I if five or more pixels in its 3-by-3 neighborhood are Figure 4(i) shows the pre-processed abdominal CT image.
Is, otherwise the pixel is set to O. The segmentation result of the improved algorithm is shown in
E. Petformance analysis Figure 4(ii). Thus the improved region growing algorithm
preceded by non linear mapping was successful in segmenting
The results are analyzed using different performance
the liver region. However the segmented image is composed
measures. To prove the accuracy of liver segmentation, the
of holes and hence the majority filling algorithm was
segmentation result of the improved region growing algorithm
introduced. The required smooth texture of the segmented
is compared to the ground truth. Performance analysis is
liver region was then obtained. Figure 4(iii) shows the post­
carried out by making use of three statistical measures and
processed image and the corresponding ground truth is shown
four similarity measures.
in Figure 4(iv).

1563
(a) (b) (c) (d)
Figure 3 (a) Raw abdominal CT image (b) Image after applying Gaussian high boost filter (c) Image after thresholding and (d) Image after non linear mapping

b(i) b(ii) b(iii) b(iv)


Figure 4 (i) Pre-processed abdominal CT image (ii) Segmented liver using the improved algorithm (iii) Post-processed image and (iv) Ground truth

The traditional region growing algorithm was applied and segmentation and the corresponding results were observed. It
the results were observed. The pre-processed CT image is was found that the results obtained by applying the K-means
shown in Figure Sea). Figure S(b) shows the segmentation clustering algorithm were different for varying number of
result of the traditional region growing algorithm. The centroids. Figure 6(a) shows the pre-processed CT image.
segmentation result of the proposed method is shown in Figure Figure 6(b-d) show the segmentation results on applying the
S(c). It was found that the traditional region growing K-means clustering algorithm for k=2, 3 and 4. The statistical
algorithm failed to segment the liver from the abdominal CT measure accuracy is graphically represented in Figure 7. The
image. Further the K-means clustering algorithm was also similarity measures are graphically represented in Figure 8.
applied to the pre- in the dataset for liver

method

Figure 6. (a) Pre-processed CT image (b) K-means clustering output for k=2 (c) K-means clustering output for k=3 and (d) K-means clustering output for k=4

1564
100

95
�>.
90
u
'" 85
Bu 80
«
75

�: ; "'; ;(., �� ,,'j ,,� ,,; ,,� ,, ' .{> ....')... V ;; � � b'"

_Clustering
���������������/ _Traditional Region Growing
• Improved Region Growing

Figure 7. Graphical representation of the statistical measure Accuracy for the K-means clustering method, the Traditional region growing method and the
Improved region growing method

Dice similarity measure Sakal and Sneath - I similarity measure


1
1

1 1 1 1 1 1 1 1 1 1 1 1 11111111111111111111111
0.8
0.6
0.4
0.2
0
���;p)���7��v;���',f��, � , � � t,��&
..:..:� v'::': ..:s ...."<2'... ..:::;..:4 V ..:..� v� .....,� ..;:;.�e v..:..� V":"'& ...:.:-:.(l; ..:...�� '-� v:-<i;

• K-Illean s clustering • Traditional .... egion g....o win g • K-means clustering • T ra ditio llClI region growing
• hnllroved region growing
• hl1l>roved region growing

(a)

Sokal and Sneath - II similarity measure Tanimoto similarity measure

1 1

II�II,IIIIIII" I" IIIIIIIIIII,IIIIIII " IIIIIII


0.8
0.95
0.6

0.4
0.9
0.2
0.35 o

���;�;6� 6 7 6�P;���5( t(y,t�1"t, ���


..:.;.'::': '>� ..:$' .::J. .$ 0..Ji....:;tfl...:;,.� v..!>.� v4e .......,:... e ...:;:�e �
v 0'::': 0'::':

• K-tneans clustering • Traditiol1al .... egion g....owing • t<-llleanS clusteri n g • T rad i tional region growing

• Irnl>roved .-egion growing • Imlu'"oved region groNing

(c) (d)
Figure 8. Graphical Representation of the similarity measures (a) Dice similarity (b) Sokal and Sneath-I similarity (c) Sokal and Sneath -II similarity and
(d) Tanimoto similarity for the k-means clustering method,the traditional region growing method and the improved region growing method.

The Table 1 shows the various performance measures of the similarity coefficient, Sokal and Sneath-I similarity
improved region growing algorithm in comparison with the coefficient, Sokal and Sneath-II similarity coefficient and the
traditional region growing algorithm and the k-means Tanimoto similarity coefficient. The average Dice similarity
clustering algorithm. It was observed that the segmentation of coefficient was found to be 0.86, the average Sokal and
the liver using the improved region growing algorithm in Sneath-I similarity coefficient was 0.72, Sokal and Sneath-II
comparison with the ground truth gives very high values of similarity coefficient was 0.98 and the Tanimoto similarity
accuracy, specificity and sensitivity. The segmentation result coefficient was 0.84 for the considered images.
is also validated by using similarity measures such as the Dice
Table I Performance measures 0f the Improved regIOn growlOg algont m 10 companson WI t e tra ItIona regIOn growlOg and the k-means c ustenng 19ont m
Algorithm A vera e Performance measures
Accuracy Specificity Sensitivity Dice Sokal and Sokal and Tanimoto
coefficient Sneath-I Sneath-II coefficient
coefficient coefficient
(%) (Normalized value)
Improved region growing 97.04 97.07 96.90 0.865 0.728 0.985 0.840
Traditional region growing 86. 63 84.80 96.82 0. 652 0.329 0.921 0.498
K-means clustering 87.52 88.88 85.23 0. 648 0.341 0.933 0.505

1565
The structural area is calculated from the segmented Iiver The proposed algorithm gave a high correlation coefficient of
image. The correlation plots are drawn between the area of the 0.7727 in comparison to the traditional region growing
segmented liver and the ground truth area. The correlation algorithm which gave a low correlation coefficient of -0.4216
coefficients are calculated from these areas. Figure 9 shows and the k-means clustering algorithm which yielded a
the correlation plot for the improved region growing method. correlation coefficient of -0.2434.

......... ,;;

--�����_�_�..---l
33k -+-
3011 31k 32k 33k 34k 35k 36k 37k 38k 3'

Groundtruth Area

Figure 9. Correlation plot of segmented liver area and ground truth area using improved region growing method

V. CONCLUSION
In this work, the segmentation of the liver from the
abdominal CT image was carried out. The segmentation [6] XU Vi, Liu Zhong and Ji Lin, "Liver CT image segmentation by Local
performed using the proposed improved region growing Entropy method", IEEE International Conference on Computer
algorithm resulted in high accuracy, specificity and sensitivity Application and System Modelling,volume II, pp.591-594,2010

as they matched well with the ground truth. A variety of


[7] Vue Huang, Jiakun Wang, Jianjun Sun and Lianfen Huang, "Liver
similarity coefficients also yielded a high value which segmentation in CT images based on Region Scalable Fitting model",
validates the proposed algorithm. Thus the improved region IEEE International Conference on Anti-Counterfeiting, Security and
growing algorithm can be used to accurately segment the liver Identification, pp.I-4, 2013
region from the abdominal CT images. The proposed
[S] Kyung Hoon Hwang and Jin Woo Jung, "Segmentation of the liver in
procedure seems to provide reliable information for diagnosis, CT images using Pseudocolorization and Labeling methods", IEEE
development of treatment plans and thus can play an important International Conference on Fuzzy Systems, pp.2170-2174,2009
role in c1inical interventions.
[9] Nader H. Abdel-massieh, Mohiy M. Hadhoud and Kamel A. Moustafa,
"A fully automatic and efficient technique for liver segmentation from
abdominal CT images", 71h International Conference on Informatics and
VI. REFERENCES
Systems, pp.I-S, 20I0
[I] Xiaoqi Lu, Jianshuai Wu, Xiaoying Ren, Baohua Zhang and Yinhui Li,
"The study and application of the region growing algorithm for liver [10] Walaa.H.Elmasry, Nashwa EI Bendary and Aboul Ella Hassanien,
segmentation",OPTIK 125,pp.2142-2147,2014 "Graph Partitioning based automatic segmentation approach for CT scan
liver images", Federal Conference on Computer Science and
[2] Shraddha Sangewar, Atish.A.Peshattiwar, Vilas Alagdevee and Rupali Information Systems,pp. l S3-IS6,2012
Balpande, "Liver segmentation of Ct scan images using K-means
algorithm", IEEE International Conference on Advanced Electronic [II] S.Priyadarshini and D.Selvathi, "Survey on segmentation of liver from
Systems, pp.6-9, 2013 CT images", IEEE International Conference on Advanced
Communication Control and Computing Technologies, pp.234-23S,
[3] Suhuai Luo, Jesse.S.Lin, Stephen K.Chalup and Guoyu Kian, "A liver 2012
segmentation algorithm based on wavelets and machine learning",IEEE
International Conference on Computational Intelligence and Natural [12] Xing Zhang, Tian Jie, Kexin Deng, Yongfang Wu and Xiuli Li,
Computing,pp.122-125,2009 "Automatic liver segmentation using a statistical shape model with
optimal surface detection", IEEE Transactions on Biomedical
[4] Hans Bhurkhardt and Wan Nural Jawahir WanYussof, "3D volumetric Engineering, vol.57, pp.2622-2626,2010
CT liver segmentation using hybrid segmentation techniques", IEEE
International Conference on Soft Computing and Pattern Recognition, [13] Gang Chen, Lixu Gu, Lijun Qian and Jianrong Xu, "An improved level
pp.404-40S,2009 set for liver segmentation and perfusion analysis",IEEE Transactions on
Information Technology in Biomedicine,vol. 13,pp.93-104,2009
[5] Xiao Song, Ming Cheng, Boliang Wang, Shaohui Huang and Xiaoyang
Huang, "Automatic liver segmentation from CT images using adaptive
fast marching method", IEEE International Conference on Image and
Graphics,pp.S97-900,2013

1566

Вам также может понравиться