A study of region based segmentation methods for

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 02 Issue: 12 | Dec-2013, Available @ http://www.ijret.org 421
A STUDY OF REGION BASED SEGMENTATION METHODS FOR
MAMMOGRAMS
Lisha Sara Varughese1
, Anitha J2
1
PG Scholar, 2
Asst.Professor, Computer Science and Engineering, Karunya University, Tamilnadu, India,
lishasaravarughese@karunya.edu.in, anitha_j@karunya.edu
Abstract
Breast Cancer is one of the most common diseases that are found in women. The number of women getting affected by cancer is
increasing year by year. Detecting cancer in the late stages, leads to very complicated surgeries and the chance of death is very high.
Early detection of Breast Cancer helps in less complicated procedures and early recovery. Many tests have been found so as to detect
cancer. Some of these tests are mammography; ultrasound etc.Mammography is a method that helps in early detection of Breast
cancer. But finding the mass and its spread from a mammographic image is very difficult. Expert radiologists are needed for accurate
reading of a mammogram. Researchers have been working for years for algorithms that help for easy detection and segmentation of
breast masses. Feature extraction and classification have also been done extensively so that the studied cases can be compared to
diagnose the new cases. Segmentation of cancerous mass regions from the breast tissues is a difficult process. Many algorithms have
been proposed for this. Some of these algorithms are region growing, watershed segmentation, clustering etc. Region Growing
Method is based on two major factors which is the seed point selection and then the stopping criteria. Watershed Segmentation on the
other hand is based on the basic geographical concept of watersheds and catchment basins, and uses a technique called as flooding. A
study of these two major region based methods such as Region Growing and Watershed Segmentation are compared and detailed in
this paper.
Keywords: Mammography, Mass Detection, Segmentation, Region growing, and Watershed Segmentation.
----------------------------------------------------------------------***------------------------------------------------------------------------
1. INTRODUCTION
Cancer is one of the most dreaded diseases and it has been
studied for years. The major cause of cancer has not been
found yet. Among the various types of cancers Breast Cancer
is a common type and is mostly found in women. A new
cancer is diagnosed every 2 minutes [1]. Late detection of
breast cancer causes the cancer cells to spread to other body
parts and organs. This will result in complicated surgeries and
also increase the chance of death. Because of these reasons
women after the age of 50 are advised by doctors to conduct
tests to detect cancer at its early stage. Mammography is one
such test that helps to detect cancer at its early stage.
Many CAD systems are available for the detection of the
breast cancer [2]. Digital Mammography helps for easy
diagnosis and several studies have been done in this area to
make the diagnosis of cancer by the expert radiologists easier.
Many algorithms have been proposed and some of them are
Region Based methods, Watershed approach, clustering,
thresholding etc
This study is performed of Region growing and Watershed
Approaches. These methods are studied based on the
performance measures used, steps followed and the number of
images the method was tested on.
2. REGION BASED METHODS
Region based methods works on a concept that the cancerous
mammogram region has some feature that is common. The
algorithms that are region based usually starts with one or
many seed points and then they are expanded. For clarity,
Region based methods can be easily classified as region
growing and split-merge method. Watershed method can be
called as a type of region growing method.
Q. Abbas et. al,2012 [3] introduced a 4 stage automatic mass
segmentation method. In the first stage of this method a
dynamic improvement of contrast in the ROI was done.
CLAHE was a method that was used in contrast enhancement.
In the second stage noise was reduced. Template matching to
determine the noise pixels and the gradient magnitude of these
pixels were calculated. Then these noise pixels were replaced.
This was followed by the third stage were a multiscale
decomposition was done on the image using steerable-pyramid
transform. Then the multiscale feature fusion of each sub
image was found. Then the candidate points were detected
from the prior and posterior probabilities based on the
multiscale feature fusion. In the fourth stage the lesion area
was characterized using maximum a-posteriori and then a
smooth mass boundary was drawn using Bezier spline curve.
This method showed effective segmentation result for
circumscribed, spiculated, ill-defined, microlobulated,
obscured and mixed masses.

__________________________________________________________________________________________
2.1 Region Growing Method
Region growing is a major type of Region Based segmentation
method. Region Growing has been illustrated in the Fig. 1.
Fig- 1: Region Growing Illustrated
The starred circle represents the initial seed points. The
immediate eight neighbors are illustrated using white circles.
Each of these pixels is added into the segmenting region if
they satisfy the thresholding condition. The grey circles show
the pixels that have satisfied the criteria. The best of these
pixels are then selected as the next seed point.
S. Meenalosini et.al, 2012 [4] proposed a method that
segmented the mass in mammograms using region growing
method. In this method pre-processing was done using
median filter, morphological operations and thresholding
methods. Segmentation was done in four steps. First contrast
enhancement was done using histogram utilization. Secondly,
alarm pixel generation was done. In this step histogram and
accumulated histogram was computed and the location of
peaks in the histogram was found out. From this the candidate
alarm pixel was calculated and from the selected candidates
the alarm pixel was generated based on a specified condition.
This alarm pixel was used in the third step which performed
region growing. The alarm pixels acted as the initial seed point
for region growing. This seed point was expanded until every
pixel had been allocated. Finally, in the fourth step Gabor
filter was applied on the image. A total of 20 Gabor filtered
images were obtained using 20 Gabor filters which was
distributed along five frequency bands. This method was
designed such that it would detect masses automatically
without any human interaction.
K. Yuvaraj et.al, 2013 [5] suggested an automatic technique of
the mass segmentation using region growing. The
mammographic image was selected and a mask was defined.
Statistical features such as mean, dissimilarity, sum average,
sum variance and auto correlation was extracted. These
features were compared with the Haralick features that were
initially extracted. If the mass features did not match then the
mask was shifted and features extracted again. This was
followed by the thresholding step were a specific threshold
value was selected for all kind of masses. Region growing was
finally performed on this image. The initial pixels of the mask
were taken as the seed points and this favored the automatic
seed selection. At the end of this step a binary segmented
image was formed. This image was scanned pixel by pixel and
each pixel was assigned with the corresponding pixel value
from the preprocessed image when the pixel value was 1.
When the pixel value was 0 it was kept 0. The results obtained
by this method were compared with the segmentation done by
an expert and it was found that the method produces good
results.
T.Berber et.al, 2013 [6] has worked on seeded region growing
algorithm and has introduced an adaptive thresholding
technique. Seeded Region growing is a type of Region rowing
where the threshold value for every mammographic image
was the same such that it caused over or under segmentation.
This method had 4 steps. First an ROI was selected which was
then made to pass through a series of pre-processing. This was
then followed by Otsu segmentation so as to find the split
point. This split point was taken as the threshold value. Thus
for every mammogram that was considered a different
threshold value was generated. The initial seed point selected
was the brightest point in the mass. The stopping criterion of
this method was that the maximum size of the mass was
calculated and the region growing was stopped based on this
criterion. This method resulted in very satisfying findings.
P. Görgel et al., 2013 [7] also proposed a region growing
based method. In this method first an enhancement of the
image was done using homomorphic filtering. Then
segmentation was done using Local Seed Region Growing
algorithm with some of the rules of Seeded Region Growing
algorithm. But the resultant ROI could contain false positives;
hence a spherical wavelet transform is applied on the ROI. A
5- level spherical wavelet transform was applied for better
performance. Two classifications were done. First
classification determined the mass area and the second
classification determined the malignancy of the mass.
2.2 Watershed Segmentation
Watershed segmentation can be understood based on the
method suggested by L. Vincent et.al. 1991 [8]. Fig -2 shows
the basic concepts of watershed, catchment basins and minima
as explained in [8].
Initial seed
point
Neighboring
pixels
Pixels in
segmented
region
Current
seed point
Neighbor

__________________________________________________________________________________________
Fig - 2: Basics Of Watershed Segmentation Illustrated
J. Herredsvela et.al, [9] showed a method for detection of
breast masses using watershed segmentation algorithm. This
method has basically two steps. In the initial step the image
was segmented using morphological watersheds. Then these
segmented regions were classified using sparse representation
of image blocks. Before applying watershed method the image
was filtered using alternating open/close sequential filters. The
image after watershed method had the suspicious mass region.
There may be a lot of false positives in these images. These
false positives were reduced using sparse representation using
learned dictionaries. Classification of these suspicious
segmented masses was done. This was done by creating a
training set using a few mammogram images. Using this
training set the other images were tested and classified. This
was called the testing phase. The error images were filtered
using Gaussian low pass filter.
J. Sharma et.al, 2011 [10] proposed a simple segmentation
technique using watershed method. In the pre-processing stage
of this method Sobel filter mask is applied to perform
convolution at 12 different orientations. Each of these
convolutions produces Curvilinear Structures (CLS). This is
followed by thresholding. The CLS is then suppressed. The
image gradient calculation is done using the Canny’s Edge
Detection. This is based on non-maximal suppression and
hysteris thresholding. Watershed segmentation based on the
Vincent-Soille approach is carried out. The image thus
obtained was ORed with the minima image obtained as the
mask image. This was then imposed on the gradient image.
The final step of this method was the segmentation phase
which is based in the catchment basin technique. Flooding is
done and flooding ends when the water reaches the highest
level. At these points virtual barriers were built which will
prevent the merging of water. The image is divided into
several regions. Each region corresponds to the region of
interest.
Wei-Yen Hsu, 2012, [11] proposed a new automatic approach
that segments the breast mass tumor using watershed method
and using the prior given information. These images were then
stored by compressing the image. Canny’s edge detection
technique was used because of its localization and
thresholding property. The Gaussian scaling parameter of
Canny’s Edge detector decides the fitness of the edge that was
detected. Edges were detected by adjusting to three criteria
which are localization, detection and response. An improved
watershed transform was performed which was based in the
previous information. This was used foe tumor edge detection.
This method can be applied to various applications based on
the amount of information that is available. Vector
Quantization and Hopfield neural network idea was also
applied. Hopfield neural network is one of the best methods
for solving optimization problems.
3. DISCUSSION
3.1 Steps Performed
There are various steps that are followed for segmenting mass
from the breast region. Some methods first pre-process the
image using various filters such as median, mean, Laplacian,
Gaussian, CLAHE etc. This pre-processing may be done
before or after the mass detection. Some methods do not
perform mass detection and define ROI. In such cases the
whole breast image is considered. Segmentation can be done
using various methods such as region growing, watershed or
clustering. Some methods also go forward to extract features
from the mass detected and this will help in the diagnosis of
the severity of the cancer mass detected. Compression of the
images help is efficient storage of the images, such that
memory space is not wasted. Table 1 shows the review of the
various steps that are followed in the various methods of this
paper.
Table -1: Review of the steps followed in various methods
Method Pre-processing Mass Detection Segmentation Feature Extraction Compression
Abbas Q. et. al,2012
Meenalosini S. et.al,2012
Yuvaraj K. et.al,2013
Herredsvela J. et.al,
Sharma J. et.al, 2011
Wei-Yen Hsu, 2012
T. Berber et.al, 2013
P. Gorgel.et.al,2013

__________________________________________________________________________________________
3.2 Database Used
Various databases are available online that provide
mammogram images. Some of them are MIAS [12], DDSM
[13], DEMS [14] etc. MIAS is Mammography Image Analysis
Society which is a database that contains a total of 322 images
which are of size 1024x1024. DDSM is Digital Database for
Screening Mammography which has about 2620 cases. DEMS
is Dokuz Eylul Mammography Set which contains 485 cases.
Table-2 shows a review of the databases that is used in the
methods discussed and the cases considered for their study.
The blank space in the table indicates the lack of specific
information.
3.3 Performance Metrics Used
Every method can me analyzed for their performance based on
measurement of their performances. This measurement may or
may not be quantitative. The performance measured is the
accuracy of the mass detected or the rate of true positives and
false positives. Performance of the particular method can also
be measured in terms of visualizing the final image and the
reference image or manually segmented image. Table-3 shows
the review of the various performance measures that are used
in the various methods discussed.
Table -2: Review of the database used in various methods
Methods Database Name
Total
number of
images
Benign Malignant
Abbas Q. et. al,2012 MIAS/DDSM 480 311 169
Meenalosini S.
et.al,2012
MIAS 250 125 125
Yuvaraj K. et.al,2013 MIAS 22 - -
Herredsvela J. et.al, MIAS 16 - -
Sharma J. et.al, 2011 - - - -
Wei-Yen Hsu, 2012 MIAS - - -
T. Berber et.al, 2013 DEMS 260 - -
P. Gorgel.et.al,2013 MIAS - - -
Table -3: Review of the metrics used in various methodsl
Method Visualization Quantitative Measure
Abbas Q. et. al,2012
Hausdroff distance, Area overlap measure,
Combined equal importance
Meenalosini S. et.al,2012 -
Yuvaraj K. et.al,2013 Mean calculation
Herredsvela J. et.al, True positive and false positive rates
Sharma J. et.al, 2011 -
Wei-Yen Hsu, 2012 Jaccard similarity index, Dice Index
T. Berber et.al, 2013 Haousdroff Distance, Yasnoff Measure etc.
P. Gorgel.et.al,2013 Sensitivity, Specificity, Accuracy
CONCLUSIONS
This paper reviews on Region Growing Method and
Watershed segmentation methods. Each of these methods have
been discussed based on the steps performed, the number of
images they have tested and the method that is used to
calculated and analyze the performance of each of these
measures. For Region Growing method the seed point and the
stopping criteria should be known. But in the case of
Watershed segmentation the pixel intensities should be known
and the concept is wrapped around the geographical watershed
and catchment basins.

__________________________________________________________________________________________
REFERENCES
[1]. http://bcaction.org/wp-content/uploads/2012/12/The-
Facts-and-Nothing-but-the-Facts.pdf
[2] H. Fujita, et al., State-of-the-art of computer-aided
detection/diagnosis (CAD), Medical Biometrics 6165
(2010) 296–305
[3] Q. Abbas et.al., Breast mass segmentation using
region-based and edge based methods in a 4-stage
multiscale system, Biomedical Signal Processing and
Control 8(2013) 204-214.
[4] Meenalosinin S et.al., Segmentation Of Cancer Cells In
Mammogram using Region Growing Method and
Gabor Features,International Journal of Engineering
Research and Applications, Vol 2, Issue 2, Mar-Apr
2012, pp. 1055-1062.
[5] K. Yuvraj et.al., Automatic Mammographic Mass
Segmentation based on Region Growing technique, 3rd
International Conference on Electronics, Biomedical
Engineering and its Applications, April 29-30,2013
Singapore .
[6] T. Berber et.al Breast mass contour segmentation
algorithm in digital mammograms, Computer Methods
and Programs in Bio-Medicine 110 (2013) 150-159
[7] P. Görgel et al. Mammographical mass detection and
Classification using Local Region Growing – Spherical
Wavelet Transform (LSRG-SWT) hybrid scheme,
Computers In Biology and Medicine 43 (2013) 765-
774
[8] L.Vincent et.al. , Watersheds in Digital Spaces : An
Efficient Algorithm Based on Immersion Simulation,
IEEE Transactions On Pattern Analysis and Machine
Intelligence, Vol.13, No. 6, June 1991.
[9]. J.Herredsvela et. al., Detection Of Masses In
Mammograms By Watershed Segmentation And
Sparse Representations Using Learned Dictionaries.
[10] J. Sharma et.al., Mammogram Image Segmentation
using Watershed, International Journal of Information
Technology and Knowledge Management, July-
December 2011, Volume 4, No.2,pp. 423-425.
[11] Wei-Yen Hsu, Improved watershed transform for
tumor segmentation: Application to mammogram
image compression, Expert Systems with Applications
39 (2012) 3950–3955.
[12] MIAS Database:
http://peipa.essex.ac.uk/info/mias.html
[13]. DDSM Database:
http://marathon.csee.usf.edu/Mammography/Database.
html
[14]. DEMS Database:
http://demir.cs.deu.edu.tr/index.php/downloads?
BIOGRAPHIES
Lisha Sara Varughese Received the B.Tech in
Computer Science and Engineering from
Mahatma Gandhi University, Kerala in 2012.
She is currently pursuing here final year in
M.Tech in Computer Science and Engineering
in Karunya University, Coimbatore. Her area of interest is
image segmentation and processing.
Mrs. J. Anitha completed her B.E. in
Information Technology with First class
distinction in 2004 from Manonmaniam
Sundarnar University, India. She completed
M.E. in Computer Science with first class
distinction in 2006 from Manonmaniam Sundarnar University.
She started her teaching career in Noorul Islam college of
Engineering from March 2006. Currently she is working as an
Assistant Professor in Karunya University from January 2007.
Her area of interest is image segmentation, image and video
compression.

A study of region based segmentation methods for

More Related Content

What's hot (16)

Viewers also liked (20)

Similar to A study of region based segmentation methods for (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

A study of region based segmentation methods for