Method to diagnosis and classify breast cancer by using Wavelet
packet & Neural Network.
Nandkumar C.Jambhale1, Prof.Ajay.S.Wadhawe2
Electronics & telecommunication department
Shri shivaji institute of engg. Management studies,Parbhani,India
Email: jambhalenandkumarc10@gmail.com,ajaywadhawe@rediffmail.com
Abstract— This paper presents a method to find and
classify the breast cancer by using wavelet packet and
artificial neural networks (ANN). The micro
calcifications present in the mammogram contain high
frequency component than other part of mammogram.
By using wavelet packet decompose the mammogram
into different frequency spectrum band. From the
frequency spectrum band select only high frequency
spectrum band and remove low frequency spectrum
band. That means remaining spectrum band contains
only high frequency component that is nothing but the
micro calcifications in the Mammogram. This result is
applied to the artificial neural networks (ANN) for the
classification of the breast cancer. The required database
(Mammograms) collected from the Swami Vivekanand
Hospital Latur (MH.INDIA).Presented system is checked
by radiologist from the same hospital.
Keywords- breast cancer;microcalcification;wavelet packet;
artificial neural networks (ANN)
I. INTRODUCTION
Cancer is a group of disease that causes cells in the body
to change & grow out of control. There are different types of
cancer such as lung, stomach, liver, colorectal, cervical,
breast cancer etc. In the whole word, every year few lakh of
people died because of cancer .In all the cancer type’s breast
cancer is the second leading cause of death in women. In US
1 in 8 women will be diagnosed with breast cancer in their
lifetime [1]. Breast cancer begins in the breast tissue that is
made up of glands for milk production, called lobules and
duct that connect the lobules to the nipple. In India Almost
90% patient presenting breast cancer when symptomatic in
stage II-B, III that means tumor spread to surrounding tissue
or nearby lymph node. Mortality rate of breast cancer are
almost twice in INDIA than US [2]. So that it is very
important to diagnosis patient in early stage of breast cancer.
There are number of tools available to diagnose the breast
cancer such as Mammography machine, Ultrasound,
Pathology, CT scan (3D), and MRI Scan. Before clinical
symptom appears to find out the early stage breast cancer,
prefer mammography machines. So the screening
mammogram is the best tool available for detecting
cancerous lesions .Since about half of cancers detected by
mammography correspond to clustered micro calcifications.
These lesions are one of the mammographic hallmarks of
early breast cancer [1].
Usually, the shape and arrangement of micro
calcifications help the radiologist to judge the likelihood of
cancer being present. Malignant calcifications are typically
very numerous, clustered, small, dot-like or elongated,
variable in size, shape and density. Benign calcifications are
generally larger, more rounded, smaller in number, more
diffusely distributed, and more homogeneous in size and
shape. However, because of the small size of micro
calcifications, the characterization of benign and malignant
lesions represents a very complex problem even for an
experienced radiologist. Moreover, there are many cases in
which the structure of malignant micro calcifications is not
very different from that of benign ones. These perceptual
problems result in screening errors that lead either to missed
malignant cases or more often to unnecessary biopsies. It
has been reported that only 17% of calcifications requiring
biopsy are cancerous [1]. Computer based diagnosis system
can help to reduce the number of false positives and
therefore the number of unnecessary biopsies.
In the literature, several techniques have been proposed
to detect the presence of micro- calcifications using various
methodologies. Concerning image segmentation and
specification of regions of interest (ROIs), several methods
have been proposed such as classical image filtering and
local thresholding [6]. Stochastic fractal models [9],
wavelet analysis [5, 10, 11]. Furthermore, various
classification methodologies have been reported for the
characterization of ROI such as, fuzzy logic systems
[4].Nevertheless, the most work reported in the literature
employs neural Networks for cluster characterization
[3,5,7,13].
In this study, we present system, aiding radiologist for
breast cancer diagnosis and identification of micro
calcification clusters in digitized mammographic images.
As the micro calcifications correspond to high-frequency
components of the image spectrum, detection of micro
calcifications is achieved by decomposing the
mammograms into different frequency sub bands,
suppressing the low-frequency sub band, and finally,
reconstructing the mammogram from the sub bands
containing only high frequencies. The wavelet transform
often fails to accurately capture high-frequency
information, especially at low bit rates where such
information is lost in quantization noise. Coifman et al.
developed a technique called wavelet packets that is better
able to represent high frequency information [13].To
achieve the best result we employed different types of
wavelet packets such as Daubichies, Symlet, Coiflet,
Biorthogonal. We use these results as inputs of neural
network for classification. The neural network contains one
input, two hidden and one output layers. A neural network
is a set of connected input/output units where each
connection has a weight associated with it. During the
learning phase, the network learns by adjusting the weights
so as to be able to predict the correct class of the input
samples. The In the false-positive reduction (fpr) step we
try to separate false signals from micro calcifications by
using a classifier based on a Support Vector Machine (SVM)
algorithm performs learning on a multi-layer feed-forward
neural network[12].
Fig.2 Decimated analysis wavelet filter bank
II. MATERIAL AND METHODS
A. Pre-processing and segmentation module
In a typical mammogram several different areas are
present such as the image background, the tissue area, and
informative marks. At the beginning of pre-processing it is
necessary to locate the breast region. For this reason we
apply a skin-line segmentation procedure by setting equal to
zero the image pixels with intensity less than 20 (for 0– 255
Gray levels). Most of those pixels belong to the background
area, although a small number exist belonging to the tissue
area close to the breast surface. The results are shown in
Figure 1. As we know mammograms tend to be low
radiation dose images which contain quantum noise called
‘mottle’, enhancing micro calcifications also enhances
mottle or noise. For getting rid of this problem, we applied
Median filters on the mammograms and subtract the result
from the original ones then wavelet transform was applied
on these images.
(a)
(b)
Fig.3 (a) wavelet output (b) wavelet packet output
The single level expansion results in 4 "details" images
dHH, dHL, and dLH, (shorter: HH, HL, LH) covering
Independent bands in the frequency domain. The
"approximation" aLL (or LL) is a low-pass component,
which is passed to the next level of decomposition.(Figure
3). Wavelet transform often fails to accurately capture
high- frequency information, especially at low bit rates
where such information is lost in quantization noise.
Coifman et al. developed a technique called wavelet
packets that is better able to represent high-frequency
information [13].A multiple level wavelet filter bank
involves iterating the lowpass-highpass filtering and
downsampling procedure only on the output of the lowpass
branch of the previous stage. Coifman et al. formulated an
extension of the octave- band wavelet decomposition to full
tree decomposition by allowing the lowpass–highpass
filtering and downsampling procedure to be iterated also on
highpass (bandpass) branches in the tree [13]. They defined
the new basis functions, called wavelet packets. We have
employed four different types of
wavelet packets:
Daubichies, Symlet, Biorthogonal, Coiflet and two level
expansions.
C. Artificial Neural Network
Fig. 1 (a) The original mammogram (b), (c) localized
Breast Border
B. Wavelet and Wavelet packet
Multi-scale representation has proven to be useful in
many Image processing applications. Wavelet analysis is one
way to generate such representation. Wavelet transform has
been used for mammographic image compression, image
enhancement, micro calcification detection, and feature
extraction[10,11,13]. Figure2 shows the structure of wavelet
packet.
2
The algorithm uses a feed-forward back propagation
network w i t h t h r e e h i d d e n layers, consisting
of 25, 10, 25 neurons for the first, second, and third
layer respectively. During training, the features a r e
extracted from the images in which the d i a g n o s i s i s
known. From t h e r e , each stage of c a n c e r was
linked through t h e neural network t r a i n i n g with
i t s three features. After training is over, the trained
networks are stored
to be used in the algorithm.
Whenever
an image is taken
as input in the
algorithm, it is simulated with the trained net- works
and from the results a percentage can be given to
which diagnosis should be taken from the mammogram.
Matlab is a good programming
toolbox package
Provides functional software e n v i r o n m e n t for
c r e a t i n g neural network. The main goal of
this package is to provide users with a set of integrated
tools neural networks to create models of biological
and simulate them easily, without the need of
extensive coding. Neural network creation is t h e
last step in t h e proposed algorithm; There is a neural
network for each stage. The code takes the desired
picture and runs it through the neural networks one
by one, each time resulting in a score. For e a c h
network we can s u m m a r i z e t he most important
operations [3,5,7,13].
TABLE 1
UNITS FOR MAGNETIC PROPERTIES CENTIMETERS
Features related to the shape and appearance of micro calcifications
III. DATABASE
We used a mammogram database developed by Dr.Aruna
Deodhar in the Swami Vivekanand Hospital Latur
(MH.INDIA). Clinical images that collected form hospitals
mentioning that biopsy has done on the patients so we
already know the results of benign or malignancy. The
spatial resolution was set to be 1012 by 938 pixels per image
with a gray-level resolution of 10 bits per pixel. There
are 149 mammograms in the database taken from 42
different patients. There are a total of 49
microcalcifications clusters, 12 benign and 37 malignant
cases, in the database with varying sizes and visibility.
Containing microcalcification clusters with different
visibility, this database is a good representative of clinical
cases.
IV. CLASSIFICATION MODULE
In order to specify the features that will be used as inputs to
the classification system, at first 25 features are identified
and computed characterizing either an individual microcalcification (object) or a group of them in a specific ROI.
Those features fall into three categories related with the
intensity, shape and texture properties of each object the
selection of the minimum two largest micro calcifications
is made since a very small micro calcification does not have
enough pixels for reliable feature value computation.(Figure
4) shows how the classification module works.
Digital
Mammogram
Normal /Abnormal
and Benign /
Malignant
Breast
Localization
Median
Filter
Neural Network
Classifier
(number of hidden layers and hidden nodes per layer)
several networks were tested with one or two hidden layers
and different number of hidden nodes.The neural network
contains one input, two hidden and one output layers.
Wavelet
Packet
----------------------------------------------------------------------------------------
Number of microcalcifications in the cluster
Maximum size of calcifications in cluster
Standard deviation of the size of calcifications in cluster
Number of calcifications with size of 1 pixel
Sum of the area of the calcifications in each cluster
Maximum value of compactness in cluster
Average compactness in cluster
Radius of the circle that best fits the cluster
Scattering of the microcalcifications
Average gray level of the microcalcifications in cluster
Standard deviation of the mean of the microcalcification gray levels in
the cluster Maximum standard deviation of the gray levels in each
calcification Average standard deviation of the gray levels in each
calcification in cluster
Area of the cluster convex hull
The length of the cluster convex hull Neighbouring with a larger
cluster Average microcalcifications intensity
Average local microcalcification background
V. Graphical user interface
The g r a p h i c al user interface w a s created t o
provide the outmost user friendliness and ease of
use. All input to t h e p r o g r a m and i t s m a i n
a l g o r i t h m is performed through the GUI, and all
the results are produced through it as well. In this
GUI[14], the picture which will be detected is
selected, and there are many factors and values
(such as: the scaling factor, bandwidth threshold,
initial index, and the initial row) of this picture helped
to apply the processing that is explained later in this
paper in more details and the processed figures are
shown in this interface. Fig.5 shows a main view
of the GUI. In the GUI t h e r e a r e many buttons:
Feature
Extraction
Fig. 4 Micro calcification diagnosis system
It must be noted that most of the selected features
correspond to the mammographic characteristics that
radiologists examine during a diagnostic procedure such as
shape, density, size, distribution of the examined group or
individual objects. In the next step of the classification
module the selected features are fed into a neural network
classification system. The neural network t h a t i s used for
characterization
is a
feed-forward back propagation
network. In order to select a p p r o p r i a t e architecture
3
Fig.5 main view of the GUI
These are the push and radio button values that change the
implementation algorithm’s various variables
and
parameters. They can be changed manually.
Run the .m file .it will open the GUI then follow te following
steps:
1. Select the input image.
2. Select the decomposition levels.
3. Select the directions.
4. Select the feature extraction.
5. Select the Training N/A.
6. Select the Training B/M.
7. Select the Normal/Abnormal.
8. Select the Benign/Malignant.
9. Close the GUI.
When you run GUI then first select the input image then
select decomposition levels that is 2,3,4 and select
directions 2,4,8,16,32 and 64 for selection of pixels. After
that select feature extraction button, it shows you
different feature images such as output of median filter,
Output of wavelet decomposition levels and output
image with micro calcification. after train the output
image for normal, Abnormal, Benign and Malignant in
step 5and 6.last you will get the result that input image is
Normal ,Benign and Malignant on EDIT window in step
7and 8.after the result Close the GUI.
Fig.6.c. image with micro calcifications
VI. RESULT AND ANALYSIS
Following images shows the result of normal, benign,
malignant image in the MATLAB with performance
characteristics.
Fig.6_e, fig.7_e and fig.8_e show the performance
characteristics of normal, benign and malignant images
respectively. The performance of images close to our
requirement.
A. Normal Image
Fig 6.d. final result of normal image on GUI
(a)
(b)
Fig. 6. (a) original Normal image (b) output of
median Filter
Fig.6 (e) Performance Characteristic of
Normal Image
4
B. Benign Image
(a)
(b)
Fig. 7 (a) Benign image (b) output of median
Filter
Fig.7 (e) Performance Characteristic of
Benign Image
C. Malignant Image
Fig.7.c. image with micro calcifications
(a)
(b)
Fig. 8(a) malignant image (b) output of median
Filter
Fig.8.(c).Image with micro calcifications
Fig 7(d). Final result of benign image on GUI
5
correctly segmenting all of the i n p u t test images.
It provides with the proposed Algorithm a very high
level of robustness.
The results of the neural
networks have better accuracy for each breast
cancer stage. The resulting diagnosis showed great
promise for being an invaluable and dependable tool
f o r the d i a g n o s i s o f breast c a n c e r . Different
methods w e r e
seamlessly joined together a n d
meshed in a highly technical algorithm w h i c h
can be considered e f f i c i e n t and v e r y easy to use.
Thus, our work show a very large area of methods
and Techniques can be successfully merged in order
to obtain a useful result for human use.
REFERENCES
[1]
Fig 8.d. final result of malignant image on GUI
Fig.8(e) Performance Characteristic of
Malignant Image
7. Conclusion
A novel technique was presented in this paper. It
incorporates
neural
networks i n
conjunction
with advanced
image processing procedure as a
method by which breast cancer diagnosis was
performed based on mammogram
pictures
obtained. The
proposed Algorithm s h o w e d great
success in
identifying the region of interest an d
6
CancerNet, A service of the National Cancer Institute
http://cancernet.nci.nih.gov
[2] http://www.who.com
[3] P. Sajda and C. Spence. Learning Contextual Relationships in
Mammograms using a Hierarchical Pyramid Neural Network
IEEE Transactions on MedicalImaging 21 (3) (2002)
[4] B. Verma and J. Zakos, \A computer-aided diagnosis system for digital
mammograms based on fuzzy-neural and feature extraction
techniques," Information Technology in biomedicine IEEE 5,
pp.46{54,March 2001
[5] Zheng L, Chan A. An artificial intelligent system for tumor
detection in screening mammogram. IEEE Trans Med
Im2001;20(7):559-67
[6] Moti Melloul, Leo Joskowicz, Segmentation of
microcalci-fication in X-ray mammograms using
entropy thresholding technical Report, May 2002,
Hebrew University, Leibniz Center
[7] Z. R. Yang and R. G. Harrison, \Detecting false benign in
Breast cancer diagnosis," Neural Networks, IEEE 3, pp.
655{658, July 2000.
[8] Dengler J, Behrens S, Desage JF. Segmentation of microcalcifications in mammograms. IEEE Trans Med Image 1993;
12:634-42.
[9] Li H, Liu KJR, Lo SCB. Fractal modelling and segmentation for
the enhancement of microcalcifications in digital mammograms. IEEE Trans Med Imag 1997;16(6):785-98.
[10] Yoshida H, Doi K, Nishikawa RM, Giger Schmidt
RA.An improved CAD scheme using wavelet
transform for detect- Ion of clustered microcalcifications in digital mammograms. Acad Radiol
1996;3:621-7
[11] Lado MJ, Tahoces PG, Mendez AJ, Souto M, Vidal JJ. A waveletbased algorithm for detecting clustered microcalcificat- ions in
digital mammograms. Med Phys 1999;26(7):1294-305
[12] Armando Bazzani , Alessandro Bevilacqua ,Dante Bollini ,
“Automatic detection of clustered micro calcifications in
Digital mammograms using an SVM classifier”
ESANN'2000 proceedings,ISBN 2-930307-00-5, pp.
195-200.
[13] Gholamali Rezai-rad, Sepehr Jamarani “Detecting
Microcalcification Clusters in Digital Mammograms Using
Combination of Wavelet and Neural Network”, Proceedings
of the Computer Graphics, Imaging and Vision: New
Trends (CGIV’05) 0-7695-2392-7/05 $20.00 © 2005 IEEE
[14] Moh’d Rasoul A . Al-Hadidi, Mohammed Y. Al-Gawagzeh
,“Solving mammography problems of breast cancer detection
Using artificial neural networks and image processing
techniques” Indian Journal of Science and Technology Vol. 5
No. 4 (Apr 2012) ISSN: 0974- 6846