See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330466810

Comparative Evaluation of Segmentation Algorithms for Tumor Cells Detection

from Bone MR Scan Imagery

Conference Paper · October 2018

DOI: 10.1109/ICISET.2018.8745612

CITATIONS READS

4 1,986

2 authors:

Eftekhar Hossain Mohammad Rahaman

University of Central Florida Bangladesh Agricultural University
37 PUBLICATIONS 608 CITATIONS 21 PUBLICATIONS 286 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Eftekhar Hossain on 26 March 2019.

The user has requested enhancement of the downloaded file.

2018 2nd Int. Conf. on Innovations in Science, Engineering and Technology (ICISET), 27-28 October 2018, Chittagong, Bangladesh

Comparative Evaluation of Segmentation

Algorithms for Tumor Cells Detection from Bone
MR Scan Imagery
Eftekhar Hossain Mohammad Anisur Rahaman
Dept. of Electronics & Telecommunication Engineering Dept. of Electronics & Telecommunication Engineering
Chittagong University of Engineering & Technology Chittagong University of Engineering & Technology
Chittagong, Bangladesh Chittagong, Bangladesh
u1308006@student.cuet.ac.bd anisur.rahaman@cuet.ac.bd

Abstract—Bone cancer is considered to be the most dangerous a number of similarity criteria [3]. It comprises thresholding-
and often the cause of early death around the globe. Therefore, based segmentation, region-based segmentation and clustering
early detection of the bone cancer has become needed to cure techniques as all of these has some predefined criteria. In
the patient. A number of segmentation methods have been
used for bone tumor detection. This study gives a comparative medical image segmentation bone cancer detection is a chal-
assessment of the existing bone cancer segmentation methods and lenging task because bone images contain granny portions of
also proposed an object labeling algorithm for the segmentation tissues and low volume tumor which make problems of over
of bone tumor from magnetic resonance images (MRI). The or under segmentation. Bone cancer is a multifarious genetic
comparison of the existing bone tumor segmentation algorithms disorder which occurs due to various physiological factors
with the proposed one has been done on the basis of quantitative
methods like the dice similarity coefficient (DSC) and the struc- and it directly affects the bone. It produces the uncontrolled
tural similarity index measurement (SSIM). The comparative growth of the cell making demonic bone tumors and invade
evaluation found that the object labeling algorithm provides the to the adjacent parts of the body. Bone cancers are also
highest mean of DSC 96.04% and mean of SSIM 98.33% over called sarcomas [4]. Basically, the bone cancer is classified as
the other segmentation methods. either primary or secondary cancer where the actual cause of
bone cancer is not known. When the malignant bone tumors
Index Terms—Bone tumor; MRI; Object labeling algorithm;
DSC; SSIM. start rising in normal bone tissues then it is called primary
bone cancer stage. Primary bone cancer rarely arises and it
counts for much less than 1 percent of all cancers. About
I. I NTRODUCTION
2300 new cases of primary bone cancer affected patient are
Image segmentation always plays a monumental role in diagnosed in the united states each year [5]. Osteosarcoma,
cancer diagnosis. Magnetic resonance imaging (MRI) or com- chondrosarcoma, and Ewing sarcoma are the most common
puted tomography (CT) are the main schemes of highlighting types of bone cancer. In the year 2014, an estimation of
the bone cancer segment from a bone anatomy [1]. The actual bone cancer affected patient is provided by the American
meaning of segmentation is the splitting of an image into cancer society (www.cancer.org) in which shows that about
several regions and then extract the meaningful information 3020 new cases have been diagnosed and 1460 deaths are
from this regions [2]. The objective of the image segmentation expected from this patient due to bone cancers. Many different
is to facilitate the representation of an image in such a segmentation algorithms have been approached throughout the
way that it becomes simpler to explore. In medical image years for bone cancer detection. The region growing algorithm,
analysis, segmentation is widely used to find out the tumor k-means clustering and fuzzy C-means clustering integrated
portion from a type of medical imaging technique whether with k-means are the already used algorithms [6] [7] [8].
it is MRI or CT scan. This study only focuses on the bone Every algorithm has its own advantages and drawbacks and
tumor detection from MR images. Based on two elementary this paper presents a comparative study of the bone cancer
properties of image intensity values the image segmentation segmentation algorithms with the proposed algorithm. In this
algorithms are distinguishable and the properties are disconti- paper, an object labeling algorithm has been approached for
nuity and similarity. In the former category, the segmentation the segmentation of bone cancer and presents a comparison
is done by finding the changes of intensity over the image. with the other existing segmentation algorithms.
It comprises techniques such as edge detection that tries to The remaining portion of the paper is structured in the
find the sharp variation in intensity between the dissimilar following way: Section II provides a summarization of the
regions and thus segments an image. The latter one is based various bone cancer segmentation algorithm that has been
on splitting an image into regions those are uniform due to completed in the fields of bone cancer detection. Section

978-1-5386-8524-2/18/$31.00 ©2018 IEEE

IV gives a brief description of the methodology that has
been proposed for the bone tumor detection. The results and
comparison of the applied methods have been disclosed in
Section V. Finally Section VI puts an end with conclusions of
the work.

II. T UMOR S EGMENTATION ALGORITHMS

In this paper existing three bone cancer segmentation algo-
rithms have been chosen for the performance comparison with
the proposed method. The methods are bone cancer segmenta-
tion using region growing algorithm, K-Means clustering and
K-Means integrated with fuzzy C-Means algorithm.

A. Region Growing Algorithm

Region growing algorithm is a region-based image segmen-
tation method in which the background and foreground pixels Fig. 2: Flow chart of k-means clustering based segmentation
are coupled together to distinguish one or more region of
interest from the whole image [9]. The division of sub-regions
or larger regions are done on the basis of some similarity growth process. The flowchart of the region growing algorithm
criterion. The most prominent advantage of region-based seg- for bone tumor segmentation is shown in Fig. 1.
mentation is that it is comparatively easier and provide greater
noise immunity over the edge detection method. B. K-Means Clustering
The region-based segmentation is accomplished by observ- K-means clustering is a learning algorithm where the
ing three points: initial seed points or pixel, neighboring pixels method of learning is unsupervised and it is widely used in
of the seed points and the determination of the region of image segmentation. Bone cancer segmentation is also done
neighboring pixels where it belongs. The process is continued by using k-means clustering. Clustering is the technique of
until the image is partitioned into regions. In the execution of arranging the uniform pixels into a number of clusters where
this algorithm, there exist three significant aspects. The first k defined the cluster number. It is an iterative algorithm and the
and the most important one is the selection of the number of arranging is made by reducing the sum of squares of distances
initial seed pixels from which the region starts growing; the between the point and the corresponding cluster center [9]. The
second one is to define the criteria or the similarity function data is assigned to the nearest cluster center and the center is
by which adjacent pixels can comprise in the growth process measured by finding the mean of all data points in that cluster.
and the last one is to select a threshold conditions to stop the The algorithmic flow chart of the k-means algorithm for bone
cancer segmentation is shown in Fig. 2.
K-means algorithm can be resolved with the following steps:
1) For a particular cluster assignment C of the data points,
the cluster means can be computed by using (1):
X
xi
j:C(j)=k
mk = , k = 1, ..., k (1)
Nk
2) Calculate the sum of squared errors between the mean
and data points using (2):
K
X X
E(C) = Nk k xj − mk k2 (2)
k=1 C(j)=k

where, mk is the mean vector of the k th cluster and Nk

is the number of observations in k th cluster.
3) For a current set of cluster means, assign each observa-
tion as (3):

C(i) = argmin k xj − mk k2 , j = 1, ..., N (3)

Fig. 1: Flow chart of region growing algorithm 4) Continue above two steps until convergence
C. Hybrid Clustering
A hybrid clustering based segmentation is the combination
of soft and hard segmentation. Basically, k-means integrated
with fuzzy c-means clustering is referred to the hybrid cluster-
ing algorithm. The integration is done due to the fact that in
some cases k-means clustering gives false segmentation and
for this reason, after applying k-means algorithm the image
is fed for fuzzy clustering which is done by fuzzy C-mean
algorithm. Fuzzy C-mean clustering is also an unsupervised
learning algorithm and it allows one piece of a data point of
one cluster to be a partial member of another cluster. The
fuzzy c-mean algorithm is based on minimizing an objective
function and the steps are:
1) Initialize the objective function which is given by (4) as
c X
X n
J(U, c1 , c2 ..cn ) = um 2
ij dij (4)
i=1 j=1

uij is the degree of membership of xi in the cluster j,

m is any real number, xi is the ith of d-dimensional
center of the cluster, cj is the d-dimensional center of
the cluster.
2) Calculate the centroid using (5)
Pn m
j=1 uij xj
cj = Pn m (5)
j=1 uij

3) Update the membership function uij using (6)

1
uij = c   2 (6)
X dij m−1
dkj
k=1

4) The iteration is stopped when the condition (7) is

satisfied
k uk+1
ij − ukij k< δ (7)
Where δ is the termination criterion and k is the itera-
tion.
The main advantage of hybrid clustering is that it consists (a) (b)
both hard and soft clustering where the hard clustering method Fig. 3: Sample images from bone tumor dataset (a) Bone MR
assign each data point to the closest cluster from the points image (b) Ground truth image
and in soft clustering method every data point is assigned a
degree of membership, rather than inserting to just one cluster
[3]. the malignant bone MR images and their respective ground
truth images. From the top, the first two samples are from the
III. DATASET D ESCRIPTION
benign type and last two samples are from the malignant type.
In this work, two datasets of bone cancer images have been
used for the performance comparison of the segmentation
algorithms. The dataset consists of total 60 bone cancer IV. P ROPOSED A LGORITHM
affected patients magnetic resonance (MR) and their ground In this paper, an object labeling algorithm has been used for
truth images. Among the two datasets, one dataset consists of bone cancer segmentation. The block diagram of the proposed
30 benign type bone cancer MR and ground truth images and algorithm is shown in Fig. 4.
second dataset consists of 30 malignant type bone cancer MR The proposed algorithm steps are filtering, binarization,
and ground truth images. All the original and ground truth object labeling, and morphological operation. Before starting
images have been taken from different orthopedics and tumor the segmentation process the MR image is converted into the
patient database. Fig. 3 shows some samples of the benign and grayscale image to reduce the computational cost and then
 (m − 1) from the set {cc0 , cc1 , cc2 , , ccm−1 } which
have the pixel density or area value dmax .

4) By using morphological operations remove all the con-

nected components whose area values are less than the
dmax and thus get the segmented tumor part.
Fig. 4: Flow diagram of the proposed bone cancer segmenta- Fig. 5 shows the segmentation output of the bone MR image
tion algorithm using proposed algorithm.
V. R ESULTS & D ISCUSSION
applied anisotropic diffusion filter to remove the noises. The The comparison of the three existing bone tumor segmenta-
advantage that detached this filter from other conventional tion techniques with the proposed algorithm have been done in
filter is that it conserves the edges sharply by blurring the MATLAB. The result section is provided the output of all the
image. On the filtered image, otsu’s thresholding method has methods and the comparative performances of this methods
been applied which performs iterating through all the possible on the basis of the number of successfully segmented MR
threshold values and calculating a measure of spread for the images, dice similarity coefficient (DSC), structural similarity
pixel levels each side of the threshold, i.e. the pixels that index measurement (SSIM) and the computational cost that is
either fall in foreground or background. The aim is to find the taken by each of the methods.
threshold value where the sum of foreground and background The segmentation output of the bone MR images for
spreads is at its minimum and thus finds the binary image different algorithms have been shown in Fig. 8 where the
which consists of only black and white pixels. Then the object segmented tumor output by each algorithm has been labeled.
labeling algorithm has been applied on the binary image to find It is evident that the object labeling algorithm provides more
the connected components. The algorithm task is to assign accurate segmentation result than other algorithms because of
a unique label to each of the components in the image like it removes all the other unnecessary tissues and arteries from
arteries, tumor, and tissues all gets an unique level. After the segmented image. In the case of k-means algorithm five
labeling the image the tumor part is segmented by using the clusters have been used for the segmentation and in the region
following steps:
1) Let there are m number of connected components
{cc0 , ccl , cc2 , , ccm−1 } in the labeled image which
is constructed from the binarized image. Let the
pixel densities are {d0 , d1 , d2 , , dm−1 } for labels
{cc0 , cc1 , cc2 , , ccm−1 } respectively.
2) Find max {d0 , d1 , d2 , , dm−1 } from this labeled con-
nected components. Let this maximum area is dmax .
3) Search for the connected component Ik , 0 ≤ k ≤

Fig. 6: Success rate plot of the different segmentation methods

(a) Input image (b) Filtered image (c) Binary image

(d) Labeled image (e) Segmented tu- (f) Detected tumor

mor

Fig. 5: Steps of bone MR image segmentation using proposed

method Fig. 7: Time complexity of the segmentation methods
 (a) (b) (c) (d) (e)

Fig. 8: Bone MR image segmentation result: (a) Input image (b) region growing algorithm output (c) k-means algorithm output
(d) hybrid clustering output (e) proposed algorithm output

growing algorithm two different seed points have been used for TABLE II: Structural Similarity Index Measurement (SSIM)
the segmentation. Total 60 bone MR images have been taken Result of Bone Cancer Segmentation Algorithms
for the performance comparison of the different algorithms
Sl.no Object Region K-means Hybrid
with the object labeling algorithm. The first performance labeling growing clustering clustering
metric is the success rate which is measured by calculating 1 1.00 0.8828 0.9523 0.9661
the ratio of the number of successful segmented bone MR 2 1.00 0.8849 0.9511 0.9189

images over the dataset which consists of 60 MR images. 3 0.9834 0.8634 0.9622 0.9524
4 1.00 0.8934 0.9228 0.9666
The comparison plot of the segmentation success rate has
5 0.9892 0.8982 0.9627 0.9632
been shown in Fig. 6. It indicates that the proposed algorithm 6 0.9812 0.8722 0.9811 0.9411
gives the highest accuracy in bone MR image segmentation by 7 0.9764 0.8599 0.9857 0.9274
8 1.00 0.8932 0.9731 0.9226
9 1.00 0.8572 0.9882 0.9195
10 1.00 0.8918 0.9594 0.9455
TABLE I: Dice Similarity Coefficient (DSC) Result of Bone
Cancer Segmentation Algorithms

Sl.no Object
labeling
Region
growing
K-means
clustering
Hybrid
clustering
segmenting 56 MR images successfully. The region growing
1 0.9873 0.8823 0.9123 0.9366 algorithm provides lowest success rate by segmenting only 50
2 0.9648 0.8153 0.9021 0.9489 MR image over 60. The k-means and the hybrid clustering
3 0.9367 0.8733 0.9213 0.9425 algorithm gives an acceptable performance in segmentation
4 0.9822 0.8526 0.9183 0.9566 success rate. The comparison plot of computational cost for
5 0.9459 0.8519 0.9367 0.9412
the segmentation methods is shown in Fig. 7.
6 0.9772 0.8823 0.9257 0.9411
7 0.9388 0.8344 0.9519 0.9578
The dice and the structural similarity index coefficients have
8 0.9636 0.8621 0.9536 0.9423
been used for the quantitative analysis of the segmentation
9 0.9524 0.8153 0.9357 0.9555 algorithms. The dice similarity coefficient (DSC) is given by
10 0.9559 0.9021 0.9293 0.9469
DSC = 2(G ∩ S)/(G + S)
TABLE III: Comparison of Bone Cancer Segmentation Algo- [8] C. K. K. Reddy, P. Anisha, and G. Raju, “A novel approach for detecting
rithms the tumor size and bone cancer stage using region growing algorithm,” in
Computational Intelligence and Communication Networks (CICN), 2015
International Conference on. IEEE, 2015, pp. 228–233.
Object Region K-means Hybrid
labeling growing clustering clustering [9] W.-X. Kang, Q.-Q. Yang, and R.-P. Liang, “The comparative research on
Mean Mean Mean Mean
image segmentation algorithms,” in Education Technology and Computer
Science, 2009. ETCS’09. First International Workshop on, vol. 2. IEEE,
DSC 96.04% 85.71% 92.86% 94.69%
2009, pp. 703–707.
SSIM 98.33% 87.97% 95.31% 94.23%

where G is the ground truth image and S is the segmented

image. The evaluated DSC value for the different algorithms
on only 10 MR segmented images have been shown in
Table. I. The performance parameter SSIM finds the structural
similarity between the ground truth and the segmented images.
Table. II shows the measured SSI values on 10 MR images
using different segmentation algorithms.
The average result of quantitative analysis on segmentation
algorithms have been shown in Table. III where it is seen that
the mean DSC and SSIM values are higher in the case of
object labeling algorithm and thus it provides maximum best
performance in bone tumor segmentation.

VI. C ONCLUSION
This work gives a comparative evaluation of existing bone
tumor segmentation techniques on MR images. This paper also
approached an object labeling algorithm for the segmentation
of bone tumor and evaluate a performance comparison of
the proposed algorithm with other methods. The segmentation
results for each algorithm has also been presented in this paper.
From the results of the comparison, it is resolved that the
object labeling algorithm provides good performance in terms
of successful segmentation rate, dice coefficient, structural
similarity index, and time complexity. Among the existing
three methods, it is found that the region growing algorithm
exhibits the worst performance over the other two techniques.
The advantage of the proposed algorithm is that the segmented
tumor output is almost similar to the ground truth image.

R EFERENCES
[1] J. Patel and K. Doshi, “A study of segmentation methods for detection
of tumor in brain mri,” Advance in Electronic and Electric Engineering,
vol. 4, no. 3, pp. 279–284, 2014.
[2] V. Janani and P. Meena, “Image segmentation for tumor detection using
fuzzy inference system,” Int J Comput Sci Mobile Comput (IJCSMC),
vol. 2, no. 5, pp. 244–8, 2013.
[3] E. Abdel-Maksoud, M. Elmogy, and R. Al-Awadi, “Brain tumor seg-
mentation based on a hybrid clustering technique,” Egyptian Informatics
Journal, vol. 16, no. 1, pp. 71–81, 2015.
[4] V. T. Devita Jr, T. Lawrence, and S. A. Rosenberg, Cancer: Principles
& Practice of Oncology: Annual Advances in Oncology. Lippincott
Williams & Wilkins, 2012.
[5] A. C. Society, Cancer facts & figures. The American Cancer Society,
2008.
[6] K. D. Mistry and B. J. Talati, “Integrated approach for bone tumor
detection from mri scan imagery,” in Signal and Information Processing
(IConSIP), International Conference on. IEEE, 2016, pp. 1–5.
[7] M. Avula, N. P. Lakkakula, and M. P. Raja, “Bone cancer detection from
mri scan imagery using mean pixel intensity,” in Modelling Symposium
(AMS), 2014 8th Asia. IEEE, 2014, pp. 141–146.

View publication stats