Journal of Biomedical Informatics 49 (2014) 45–52

Contents lists available at ScienceDirect

Journal of Biomedical Informatics

journal homepage: www.elsevier.com/locate/yjbin

Computer-aided detection of breast cancer on mammograms: A swarm

intelligence optimized wavelet neural network approach
J. Dheeba a,⇑, N. Albert Singh b, S. Tamil Selvi c
a
Dept. of Computer Science and Engineering, Noorul Islam Centre for Higher Education, Kumaracoil, Kanyakumari District, Thuckalay, Tamil Nadu 629 180, India
b
BSNL Nagercoil, India
c
Department of Electronics and Communication Engineering, National Engineering College, Kovilpatti, India

a r t i c l e i n f o a b s t r a c t

Article history: Breast cancer is the second leading cause of cancer death in women. Accurate early detection can effec-
Received 26 June 2013 tively reduce the mortality rate caused by breast cancer. Masses and microcalcification clusters are an
Accepted 17 January 2014 important early signs of breast cancer. However, it is often difficult to distinguish abnormalities from nor-
Available online 6 February 2014
mal breast tissues because of their subtle appearance and ambiguous margins. Computer aided diagnosis
(CAD) helps the radiologist in detecting the abnormalities in an efficient way. This paper investigates a
Keywords: new classification approach for detection of breast abnormalities in digital mammograms using Particle
Mammograms
Swarm Optimized Wavelet Neural Network (PSOWNN). The proposed abnormality detection algorithm is
Breast cancer
Computer aided diagnosis
based on extracting Laws Texture Energy Measures from the mammograms and classifying the suspicious
Receiver operating characteristic regions by applying a pattern classifier. The method is applied to real clinical database of 216 mammo-
Laws grams collected from mammogram screening centers. The detection performance of the CAD system is
Particle swarm optimization analyzed using Receiver Operating Characteristic (ROC) curve. This curve indicates the trade-offs between
sensitivity and specificity that is available from a diagnostic system, and thus describes the inherent dis-
crimination capacity of the proposed system. The result shows that the area under the ROC curve of the
proposed algorithm is 0.96853 with a sensitivity 94.167% of and specificity of 92.105%.
Ó 2014 Elsevier Inc. All rights reserved.

1. Introduction Mammography has been proved to be effective in screening

asymptomatic women to reduce mortality by as much as 30%.
Breast cancer is the most frequently diagnosed cancer in wo- The American Cancer Society recommends that all women aged
men worldwide and the leading cause of cancer death among fe- above 40 undergo screening mammography once in a year. Dense
males. Breast cancer accounts for 23% of the total cancer cases breast tissue can look white or light gray on a mammogram. This
and 14% of the cancer death in both developed and developing can make mammograms harder to interpret in younger women,
countries. It is estimated that more than 1.6 million new cases of who tend to have denser breasts. Many breast conditions mimic
breast cancer occurred among women worldwide in 2010 [1,2]. the symptoms of cancer and need tests and sometimes a biopsy
In 2011, nearly 1.7 million people were told to have breast cancer; for diagnosis. False positive results occur when mammogram finds
statistics says that in USA 527 new cases of breast cancer were something that looks like cancer, but turns out to be benign (not
diagnosed per day and 110 people die of it per day. Early diagnosis cancer). Depending on the density of the breasts radiologists may
remains important for survival, particularly in low and middle in- miss up to 30% of breast cancers [3]. Even qualified radiologists
come countries where the diseases is diagnosed in late stages and find it difficult to interpret screening mammograms in large
resources are very limited. One proven way of reducing mortality numbers.
from breast cancer is the screening of asymptotic women by Two powerful indicators of cancer that are commonly used in
mammography. evaluating mammograms are known as masses and microcalcifica-
Mammography is the best screening tool that uses low dose tions. It is generally accepted that mass detection is a more chal-
X-rays to create an image of the breast to find breast cancer. lenging problem than the detection of micro-calcifications, not
only for the large variation in size and shape in which masses
⇑ Corresponding author. can appear in a mammogram but also because masses often exhi-
E-mail addresses: [email protected] (J. Dheeba), mailalbertsingh@yahoo. bit poor image contrast [4]. A true abnormality can usually be dis-
co.in (N. Albert Singh), [email protected] (S. Tamil Selvi). tinguished based on careful analysis on two different view of a

http://dx.doi.org/10.1016/j.jbi.2014.01.010
1532-0464/Ó 2014 Elsevier Inc. All rights reserved.
46 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52

mammogram namely Carniocaudal (CC) view and MedioLateral sess textural information that could bear discriminant features.
Obligue (MLO) view [5]. But, additional methods or imaging is need The Laws texture features were extracted from the mammogram
if the suspected region is not viewed on the complementary view to differentiate between abnormal and normal pixels. New artifi-
of the mammogram or if there is a dense tissue present that could cial intelligent techniques such as neural network have been used
obscure an underlying abnormality. in medical applications for discriminating the normal and abnor-
Computer-aided detection and diagnosis (CAD) can be used on mal tissues in mammograms. The thriving of artificial intelligence
the digital images to help the radiologists analyse the overall which utilizes the human experience in a more relaxed form than
images, and highlights potential areas of concern that needs closer the conventional mathematical approach has recently attracted
study. CAD can find tumours that a radiologist might not spot. more attention.
Once a CAD analysis has been done, a radiologist will do a visual Designing optimal neural network architecture is made by a hu-
check of those areas, and based on training and experience, decide man expert and it requires a tedious trial and error process. Espe-
how serious the lumps may actually be [6]. CAD will assist the cially automatic determination of artificial neural network
radiologists by serving them as a ‘‘second reader’’. The proposed parameters is the most critical task. This paper focuses mainly on
CAD system will automatically identify the areas of abnormal con- designing a CAD system based on the optimized wavelet neural
trast, calling the radiologist’s attention to suspicious regions. Com- network evaluated using Particle Swarm Optimization approach
bining mammography with CAD will improve the ability to find (PSOWNN). Optimization of WNN is carried out to improve the
cancer [7]. In many cases, the microcalcifications and the cancer classification accuracy in breast cancer detection thereby reducing
masses are hidden in the intense breast tissues especially in youn- the misclassification rate.
ger women, making both the diagnosis and detection more com-
plex and intricate [8]. While mammography has been proven to
2. Materials
be a powerful tool in the fight against breast cancer, the accurate
reading of mammograms can sometimes be difficult [9]. Even the
The proposed CAD system is based on a pattern recognition sys-
most trained radiologist can miss subtle variations in tissue that
tem which intelligently identifies the abnormal regions. CAD
might be of concern.
schemes using digital image processing techniques have the goal
A lot of researches in the area of CAD systems for breast cancer
of improving the detection performance. Typically CAD systems
and developing intelligent techniques for improving classification
are designed to provide a ‘‘second opinion’’ to aid rather than
accuracy have been conducted in last few decades [10–12]. Differ-
replacing the radiologist. Fig. 1 shows the proposed approach for
ent studies have demonstrated that Computer Aided Detection
detection of abnormality in mammograms.
(CAD) of breast cancer can improve the detection rate from 4.7%
to 19.5% compared to radiologists. Regarding classification of
abnormalities in mammogram, a number of techniques have been 2.1. Mammogram database
presented using machine learning approaches to classify samples
as normal and abnormal. Clinical mammogram database consisting of 216 images of 54
Karahaliou et al. [13] investigated multi-scale texture proper- patients were taken from mammogram screening centers. The real
ties of the tissue surrounding microcalcifications (MCs) for breast time database includes a wide spectrum of cases that are difficult
cancer diagnosis using probabilistic neural network. Kupinski and to classify by radiologists. All clinical mammograms that were col-
Giger [14] presented a radial gradient index based algorithm and lected from screening clinics were positive for presence of abnor-
a probabilistic algorithm for detecting lesions in digital mammo- malities. Mammograms were collected from 54 patients and all
grams. Sahiner et al. [15] used a Convolution Neural Network these patients have agreed to have their mammograms to be used
(CNN) classifier to classify the masses and the normal breast tissue. in research studies. For each patient 4 mammograms were taken in
Eltonsy et al. [16] presented a method based on the presence of two different views, one is the Craniocaudal (CC) and the other is
concentric layers surrounding a focal area with suspicious morpho- the Mediolateral Oblique (MLO) view. The two projections of each
logical characteristics and low relative incidence in the breast re- breast (right and left) were taken for every case. The suspicious re-
gion. Zheng et al. [17] presented a mixed feature based neural gions were identified by the automated system based on various
network for detection of microcalcification clusters in digital mam- machine learning algorithms and was reviewed by experienced
mograms. Features are computed in both the spatial and spectral radiologists. For this study a total of 216 mammograms were ta-
domain and uses spectral entropy as a decision parameter. Back- ken, all the mammograms were digitized to a resolution of
propagation with Kalman filtering (KF) is employed to allow more 290  290 Dots per Inch (DPI) which produces 24 bits/pixel. Each
computationally efficient training as required for evaluation of dif- digitized mammograms was incorporated into a 2020  2708 pixel
ferent features and input images. image (5.47 Mpixels). Screening mammography is taken on
Among existing CAD techniques, the main problem of develop- asymptotic women to detect clinically occult cancers. Table 1
ing an acceptable CAD system is inconsistent and low classification shows the summary of the different types of abnormalities in the
accuracy. In order to improve the training process and accuracy, in database. The mammograms obtained are from women with an
this paper a novel intelligent classifiers that use texture informa- age group of 20–69 years old.
tion as input to classify the normal and abnormal tissues in mam-
mograms is investigated. Moreover, the intelligent machine 2.2. Mammogram image preprocessing
learning classifiers are optimized using heuristic algorithms for
finding appropriate hidden neurons, learning rate and momentum The goal of pre-processing the image is to simplify recognition
constant during the training process. of cancers (abnormalities) without throwing away any important
This paper concentrates on developing a CAD system as an arti- information. Mammograms has breast region and is superimposed
ficial second radiologist. Texture helps to understand image con- over background structures to which analysis is not necessary. One
tent based on textural properties in images. Texture is the most way would be to restrict the analysis to Region of Interest (ROI)
important visual cue in identifying different types of homogeneous that does not contain any background. The initial preprocessing
regions and gives information about the surface property, depth is done in the digital mammogram to separate the region of inter-
and orientation [18–22]. This texture information helps to extract est (breast) and the dark background. The separation of ROI from
specific characteristics from a data. Mammographic images pos- the dark background is done using a global thresholding technique.
 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52 47

Feature Extraction
Mammogram Digitizing Pre-Processing
Acquisition Laws Texture features

Real Clinical DB

Classification using
PSOWNN

Abnormal Normal

Fig. 1. Proposed CAD system for breast cancer detection.

These kernels are chosen for finding the level, edge, spot, wave
Table 1 and ripple [24].
Summary of the clinical database used in this
work. L5 ¼ ½ 1 4 6 4 1
Class of malignancy # of Images E5 ¼ ½ 1 2 0 2 1
Circumscribed masses 44 S5 ¼ ½ 1 0 2 0 1 
Ill-defined 76
W5 ¼ ½ 1 2 0 2 1
Obscured masses 10
Spiculated masses 62 R5 ¼ ½ 1 4 6 4 1
Microcalcification 24
From this above 1-D convolution kernels 25 different two
Total 216
dimensional convolution kernels are generated by convoluting a
vertical 1-D kernel with a horizontal 1-D kernel. Similarly, 25 dif-
ferent two dimensional masks can be formed.
2 3 2 3
1 1 4 6 4 1 L5L5 E5L5 S5L5 W5L5 R5L5
Consider an input mammogram image f(x, y), having light breast 6 7
6 2 7
6 7
6 2 8 12 8 1 7 L5E5 E5E5 S5E5 W5E5 R5E5
6 7 6 7
6 0 7  ½1 4 6 4 1 ¼ 6 0 0 7
area on a dark background. The objects from the background are 6
6
7
7
6
6
0 0 0 7
7
L5S5 E5S5 S5S5 W5S5 R5S5
4 2 5 4 2 8 12 8 2 5 L5W5 E5W5 S5W5 W5W5 R5W5
separated using a threshold value T and is defined as in Eq. (1). 1 1 4 6 4 1 L5R5 E5R5 S5R5 W5R5 R5R5
Then any point (x, y) for which f(x, y) > T is called the breast area;
otherwise, the point is called the background region. The threshold The following steps will describe how texture energy measures are
is chosen by visual inspection of the image histogram [23]. identified for each pixel in the ROI of a mammogram image.
Step 1: Apply the two dimensional mask to the preprocessed

f; f ðx; yÞ > T image i.e. the ROI to get F(i, j), where F(i, j) is a set of
Iðx; yÞ ¼ ð1Þ
0; f ðx; yÞ 6 T 25 N  M features.
Step 2: To generate the LTEM at the pixel, a non-linear filter is
An intensity histogram is constructed and the local threshold applied to F(i, j). The local neighborhood of each pixel
value is chosen by statically examining the intensity values of is taken and the absolute values of the neighborhood
the local neighborhood of each pixel. The mean of the local inten- pixels are summed together. A 15  15 square matrix
sity distribution is calculated and used as a threshold value. The is taken for doing this operation to smooth over the gaps
breast area in the mammogram only covers about 30%, on average, between the texture edges and other micro-features.
of each mammogram. Based on this observation, the breast area is The non-linear filter applied is,
first segmented out in order to save processing and then further
processing is restricted to the breast area. X
7 X
7
Eðx; yÞ ¼ jFðx þ i; y þ jÞj ð2Þ
j¼7i¼7
2.3. Feature extraction
By applying Eq. (2), to the ROI image 25 texture energy mea-
The texture energy measures developed by Kenneth Ivan Laws sures for each pixel is obtained.
at the University of Southern California have been used for many
diverse applications [24]. These texture features are used to ex- Step 3: The texture features obtained from step 2 is normalized
tract Laws Texture Energy Measures (LTEM) from the ROI con- for zero-mean.
taining abnormality and normal tissue patterns. These measures
are computed by first applying small convolution kernels to the The goal of texture classification then is to produce a classifica-
ROI and then performing a windowing operation. The 2-D convo- tion map of the input image where each uniform textured region is
lution kernels for texture discrimination are generated from the identified with the texture class it belongs to.
following set of 1-D convolution kernels of length five. Here tex-
tures are considered as a quantitative measure of arrangement of 3. PSO optimized WNN
intensities in a region. The kernels are chosen in a way that they
have to average to zero. The kernels are chosen so that they are Classification has been examined in diverse fields like image
sensitive to edge points, spots, lines and a combination of these. processing, pattern recognition, medical imaging, etc. Pattern clas-
48 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52

sification is assigning an object to one of a set of classes based upon velocity. The potential solutions called particles fly through the
the features of that object. The classification process mainly de- problem space by following the current optimum particles. All of
pends on how well the discriminant features performs and the the particles have fitness values which are evaluated by the fitness
classifier chosen. Intelligent algorithms have the ability to reason function to be optimized, and have velocities which direct the
and learn in an environment of uncertainty and imprecision, soft flying of the particles. Each particle is updated after every iteration
computing, an innovative approach in constructing computation- using two values pbest and gbest. pbest is the personal best
ally intelligent systems. Soft computing consists of several com- value, which indicates the best solution achieved so far (i.e. lowest
puting paradigms including feedforward neural network which fitness value) and the global best solution achieved so far by any
has its strength in learning from examples and adapting to chang- particle in the population. In a n-dimensional search space,
ing environment [25], radial basis function [26], support vector ~
X i ¼ ðX i1 ; X i2 ; ::::::; X in Þ and ~
V i ¼ ðV i1 ; V i2 ; ::::::; V in Þ are the positions
machine [27] and wavelet neural network [28]. and velocities respectively and they are updated for the dth dimen-
Wavelet Neural Networks (WNN) is an efficient model for non- sion of the ith particle and is given by,
linear pattern recognition [29,30]. The wavelet transformation
technique is used for obtaining information from signals that are V id ðt þ 1Þ ¼ V id ðtÞ þ c1  rand1  ðpbestid  X id ðtÞÞ þ c2  rand2
aperiodic, noisy, intermittent or transient. Wavelets have the prop-
erties of both spatial and frequency domain characteristics and  ðgbestd  X id ðtÞÞ ð3Þ
hence it works well in detecting the abnormalities in mammo-
X id ðt þ 1Þ ¼ X id ðtÞ þ V id ðt þ 1Þ ð4Þ
grams. Among many artificial intelligent methods, the feedforward
ANN is the widely used statistical tool designed to diagnose path-
ological images especially of cancers and precancers. Thus, the c1 and c2 are the acceleration constants, rand1 and rand2 are the ran-
study will strengthen the foundation of ANN in CAD application dom numbers, pbestid is the individual’s personal best i.e. the local
by combining the wavelet multiscale theory and neural network best solution found so far. gbestd is the neighborhood’s best solution
and obtain a novel high performance network – wavelet neural found in the entire global community or in some neighborhood of
networks. These networks are a new powerful tool for approxima- the current particle. Vid(t): Velocity of individual at iteration t, Xid(t):
tion and deals effectively with the problems of high dimensional Position of individual at iteration t, pbestid: Best position of
model [31]. individual until iteration t, gbestd: Best position of the group until
A wavelet neural network was first introduced by Zhang and iteration t.
Benvniste [28] as a class of feedforward networks composed of Back propagation training is a gradient descent algorithm and is
wavelets. The discrete wavelet transform is used for analyzing susceptible to getting trapped to the nearest local minimum. In or-
and synthesising feedforward neural network. The wavelet net- der to find optimal network architecture for the problem under
work uses wavelet activation function and preserves the universal study, exhaustive back propagation training is done over every net-
approximation property. WNN is a feedforward neural network work configuration in the architecture space defined. Performing
with an input layer, hidden layer and an output layer. The hidden the training for larger number of times with randomized initial
layer is comprised normally of wavelets as activation function and parameters increases the chances of converging to the global min-
the output layer is comprised of linear activation function. The out- imum of the fitness function. Even if the configuration is made to
put layer of WNN represents the weighted sum of the hidden layer train large number of times still there is no guarantee of converg-
units i.e. wavelet basis function. The backpropagation learning ing to the global optimum with the backpropagation. However a
algorithm is used to update the network weights and to further best performance configuration can be achieved in the architecture
minimize the standard Mean Square Error (MSE) of the networks space defined by the optimality of the network evolved using a
approximation after network construction. Tuning a WNN is more Particle Swarm Optimized Wavelet Neural Network (PSOWNN).
important because of the following reasons, The pseudo code for PSOWNN algorithm is as follows
Learning rate parameter, which, if not set properly, can either
lead to oscillation or an indefinitely long training time.
randomly generate initial population
Momentum constant parameter is to accelerate the conver-
do
gence of error propagation algorithm. The BP is just a gradient des-
for i = 1 to population_size
cent algorithm on the error space, which can be complex and
Calculate fitness value using Eq. (6)
contain many deceiving local minima. Therefore, the BPs are most
if (F(Xi) < F(pbestt1)) then
likely gets trapped into a local minimum, making it entirely depen-
pbestt = Xi
dent on initial settings.
else
Number of hidden layers and hidden neurons, determination
pbestt = pbestt1
of the optimal number of hidden layers and hidden neurons is the
gbestt = min (gbestneighbors)
most critical task. An ANN will not be capable of classifying a com-
for d = 1 to dimensions
plex set of problems with no or few hidden neurons. In contrast, if
Velocity update using Eq. (3)
the ANN has too many neurons/layers, it eventually leads to more
Position update using Eq. (4)
complex networks and can also be highly time-consuming. The
end
optimum number of hidden nodes/layers might depend on input/
end
output vector sizes, training and test data sizes and more impor-
while maximum iterations (t) is reached
tantly the characteristics of the problem.
In an attempt to improve the classification accuracy of the WNN
classifier PSO is used to tune the initial network parameters. Parti- In the proposed method, PSOWNN is applied for evolving fully
cle Swarm Optimization (PSO) was introduced by Kennedy and connected wavelet neural network and is optimized with best net-
Eberhart [32] as a population based stochastic search and optimi- work architecture by optimizing the number of neurons in the hid-
zation process. PSO simulates the behavior of bird flocking or fish den layer, the learning rate and the momentum factor. Finding an
schooling and used it to solve the optimization problems. In the ba- optimal learning rate avoids major disruption of the direction of
sic PSO algorithm the system is initialized with a population of ran- learning when very unusual pair of training patterns is presented.
dom solutions and searches for optima by updating positions and The main advantage of using optimal momentum factor is to accel-
 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52 49

erate the convergence of error propagation algorithm. The number When the number of hidden layer neurons is increased there is
of neurons in the input layer and output layer is fixed based on the an increase in the classification accuracy but at 140 hidden layer
problem defined. neurons the performance starts to degrade. This shows the impor-
Let NI represents the number of the neurons in the input layer tance of choosing optimal network parameters to increase the clas-
and NO represents the number of the neurons in the output layer. sification performance. The optimization of classifiers was
The number of neurons in the input and output layer are fixed performed with the learning rate and the momentum factor varied
and they are same for the entire configuration in the architecture from 0 to 1 and the hidden neurons varied from 31 to 200. For this
space. The number of hidden layers in this problem is restricted training a maximum of 100 generations are performed with a pop-
and made as one. The range of the optimization process is defined ulation size N = 50 and with 500 training epochs. These patterns
by two range arrays Rmin = {Nhmin, Lrmin, Mcmin} and Rmax = {Nhmax, contain normal, benign and malignant tissues. An optimized
Lrmax, Mcmax} where Nh is the number of neurons in the hidden WNN is achieved with Nh = 116, Lr = 0.00127 and Mc = 0.9283.
layer, Lr is the learning rate and Mc is the momentum factor. Let Testing is done for all the 216 real time clinical images. Masses
f be the activation function and is given as, are characterized by the margin of the mass. The mammographic
  border between the mass and the normal tissue is useful for pre-
ypk ¼ f spk ð5Þ
dicting benign and malignant masses. Margins of a mass is de-
P
where spk ¼ j wj;k ypj þ hk ; ypk is the output of the kth neuron when a scribed a circumscribed, obscured, ill defined and spiculated. The
pattern p is fed, wj,k is the weight from the jth neuron and hk is the circumscribed masses have a well defined. Fig. 2 shows detection
bias value of the kth neuron in the hidden layer. The hidden layer of a abnormalities in mammograms using PSOWNN classifier.
uses a wavelet activation function and the output layer uses linear The real time mammograms are taken in varying intensities and
activation function. The fitness function sought for optimal training hence detecting abnormalities in mammograms is a difficult task.
is the Mean Square Error (MSE) formulated as, Fig. 2(a) shows the original image of the breast mammogram and
(b) shows the result after segmentation of ROI. Fig. 2(c) shows
XX
No
 2
MSE ¼ tPk  yp;o ð6Þ the detected abnormality. The most difficult to diagnose mass in
k
p2T k¼1 mammograms are ill defined and spiculated masses. The shape of
these masses are irregular and have ill defines and spiculated mar-
where t pk is the target (desired) output, yp;o
k
is the actual output from gins. The margins of a mass are difficult to evaluate on screening
the kth neuron in the output layer o, for the pattern p in the training mammograms and this shows the importance of texture features
set. With the framed fitness function the PSOWNN algorithm auto- in mass detection. The result shows the detection of a spiculated
matically evolve a best solution. The optimally designed WNN has mass in a mammogram of a 28 year old woman using the proposed
three-layer architecture: an input layer, hidden layer and an output classifier. Microcalcification clusters are tiny calcium deposits and
layer. The number of neurons that structures the input layer is these clusters that fail to demonstrate features characteristic of
equal to the number of feature vectors extracted (25 LTEM). The benignity have to be evaluated to determine for malignancy and
hidden layer neurons are optimally added to the WNN and are de- their exact location in the breast. A mammographically significant
fined by the wavelet activation function. The output layer contains cluster is usually considered to be 3–10 calcific particles within a
one neuron which discriminates presence of abnormal and normal volume of 1 cm2. The result shows the detection result of microcal-
tissues. The neural network architecture space is defined over a cification clusters in real clinical mammograms by considering the
multilayer perceptron with the parameters range set as Rmin and Laws texture features.
Rmax. The results also show the detected masses in a denser mammo-
gram. In the actual interpretation of mammographic abnormality
4. Experimental results detection, the texture features play a major role in defining micro-
calcification and masses. The irregular shaped clusters of masses
Clinical mammograms collected from 54 patients were used for and microcalcifications without well-defined boundaries of abnor-
evaluating the proposed PSOWNN classifier. The proposed system malities are difficult to diagnose. Masses and microcalcification
has been designed in a framework of MATLAB 7.10, which aims at represent the most challenging signs characterizing the early
developing a CAD system for breast cancer detection. After prepro- breast cancer. Masses can become visible as dense region of unsta-
cessing the clinical mammograms, 25 LTEM features are extracted ble size varying from 3 to 30 mm and properties can be character-
from each pixel of the ROI image. The data is taken from ized as circumscribed, spiculated and ill defined. The great
15  15 pixels from the ROI and is subjected to further analysis. variability of their appearance represents the main difficulty of
The training dataset contains 1064 pairs of input–output training developing a mass classification method. The mammogram images
patterns from the mammograms of real clinical database. Table 2 are irregular textures with subtle differences between benign and
shows the classification performance of Feedforward Neural Net- malignant tissues. Similarly, microcalcifications are tiny calcium
work (FFNN) and WNN classifiers by varying the hidden layer neu- deposits which range from 0.2 to 0.5 mm. These clusters are also
rons with learning rate and momentum constant set as 0.01 and difficult to identify as these malignant clusters get camouflaged
0.9 respectively (values taken for FFNN and WNN design) using in the normal breast tissues and make the detection intricate.
Laws texture features. Masses are more difficult to detect than MCs because the features
of a mass may be obscured by or be similar to those of normal
breast parenchyma. The detection results illustrates that most of
Table 2
Classification accuracy by varying the hidden layer neurons.
the abnormality clusters in the mammograms are correctly
detected.
No. of hidden neurons Classification accuracy for real clinical database (%)
FFNN WNN
80 73.47 80.42 5. Performance analysis and discussion
100 76.1 83.63
125 77.52 85.153 This section describes about the various performance measures
140 76.836 84.75
used to analyse the potential of various classification approaches
160 76.05 84.351
used for breast cancer detection [33]. For evaluation of detection
50 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52

Fig. 2. Classification results of abnormalities using PSOWNN approach. From top to bottom are the cases of circumscribed masses, Spiculated masses and microcalcification:
(a) Original Image, (b) ROI image and (c) Classified abnormalities.

performance, the number of True Positives (TP), False Positives ity that the test results from a randomly chosen diseased individ-
(FP), True negatives (TN) and False Negatives (FN) should be taken ual is more indicative of disease than that from a randomly
into consideration. A true positive is considered as detected if at chosen non-diseased individual. The overall performance of diag-
least one finding is located in the associated truth marking by nostic systems has been measured and reported in terms of classi-
the radiologists. All findings outside the truth marking by the fication accuracy which is the percentage of diagnostic decisions
radiologists are considered as false detections. In this way the true that proved to be correct. Youden’s index (J) [36] is a single statistic
positive rate can be plotted against the false positive rate. Each that captures the performance of a diagnostic test and is calculated
decision threshold results in a corresponding operating point on as, sensitivity + specificity  1.
the curve. Such a curve is referred as Receiver Operating Character- There are possibilities for an automated classifier to miss cer-
istic (ROC) curve. The ROC is the appropriate measure for the pro- tain abnormalities and may miss predict the normal cases as
posed detection system because there will be a trade-off between abnormalities, this is known as the misclassification rate. In order
the true positive rate (TPR) and the false positive rate (FPR). TP, TN,
FP and FN are the four different possible outcomes of a single pre-
diction for a two-class case with classes ‘‘1’’ (for abnormal/malig-
nant cases) and ‘‘0’’ (‘‘for normal cases’’).
Sensitivity and specificity are statistical measures of the perfor-
mance of a binary classification test. Sensitivity measures the pro-
portion of actual positives which are correctly identified when the
mammogram contains cancers tissues in it. Specificity measures
the proportion of negatives which are correctly identified when
cancer is not present in the mammogram [34,35]. The sensitivity
and specificity is given in Eqs. (7) and (8) respectively.

TP
sensitiv ity ¼ ð7Þ
ðTP þ FNÞ

TN
specificity ¼ ð8Þ
ðTN þ FPÞ
The area under the ROC curve (AUC or AZ) is a measure of how
well a parameter can distinguish between two diagnostic groups
(abnormal/normal tissues). AUC can be interpreted as the probabil- Fig. 3. A comparison of ROC curve for various classifiers.
 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52 51

Table 3 References
Performance measures for optimally tuned classifier models.

Classifiers/metrics Performance measures for real clinical database [1] Jemal A, Bray F, Center MM, Ferlay J, Ward E, Forman D. Global cancer
statistics. CA Cancer J Clin 2011;61(2):69–90.
SONN [37] DEOWNN [38] PSOWNN [2] Forouzanfar MH, Foreman KJ, Delossantos AM, et al. Breast and cervical cancer
in 187 countries between 1980 and 2010: a systematic analysis. Lancet
Sensitivity (%) 90.984 93.333 94.167
2011;378(9801):1461–84.
Specificity (%) 86.111 89.474 92.105 [3] Kolb TM, Lichy J, Newhouse JH. Comparison of the performance of screening
Accuracy (%) 89.873 92.405 93.671 mammography, physical examination, and breast US and evaluation of factors
AUC 0.91382 0.95735 0.96853 that influence them: an analysis of 27,825 patient evaluations. Radiology
Youden’s index 0.77095 0.82807 0.86272 2002;225(1):165–75.
Misclassification rate 0.10127 0.075949 0.063291 [4] Cheng HD, Cai X, Chen X, Hu L, Lou X. Computer-aided detection and
classification of microcalcifications in mammograms: a survey. Pattern Recogn
2003;36(12):2967–91.
[5] Giger ML, Karssemeijer N, Armato SG. Computer aided diagnosis in medical
to investigate the usefulness of the texture features used, analysis imaging. IEEE Trans Med Imaging 2001;20(12):1205–8.
was made for Laws texture features to determine the suitable [6] Cheng HD, Shi XJ, Min R, Hu LM, Cai XP, Du HN. Approaches for automated
detection and classification of masses in mammograms. Pattern Recogn
discriminant features.
2006;39(4):646–68.
A comparative performance analysis based on ROC curve of [7] Metz CE. ROC methodology in radiologic imaging. Invest Radiol 2001;21(9):
various classifiers is shown in Fig. 3. The result emphasizes the 720–33.
[8] Motakis Efthimios, Ivshina Anna V, Kuznetsov Vladimir A. Data-driven
potential of the PSOWNN learning algorithm to be used as a breast
approach to predict survival of cancer patients. IEEE Eng Med Biol Mag
cancer classifier. The classifier focuses the experimentation on try- 2009;28(4):58–66.
ing to improve the classification rate by focussing on initial neural- [9] Bird RE, Wallace TW, Yankaskas BC. Analysis of cancers missed at screening
network settings. Consequently, by using optimized parameter set- mammography. Radiology 1992;184(3):613–7.
[10] Baker JA, Rosen EL, Lo JY, Gimenez EI, Walsh R, Soo MS. Computer-Aided
tings for wavelet neural networks the classification accuracy is im- Detection (CAD) in screening mammography: sensitivity of commercial CAD
proved drastically. A very high classification rate was achieved for systems for detecting architectural distortion. Am J Roentgenol 2003;181(4):
the optimally tuned wavelet neural networks. 1083–8.
[11] Chan HP, Doi K, Vyborny CJ, Schmidt RA, Metz CE, Lam KL, et al. Improvement
The performance evaluation demonstrates that the result of the in radiologists detection of clustered microcalcifications on mammograms.
optimized wavelet neural network classifier is generally better The potential of computer-aided diagnosis. Invest Radiol 1990;25(10):
than that of other well known classifiers. This is consistent with 1102–10.
[12] Hindle William H. Breast care: a clinical guidebook for women’s primary
the fact that optimization is useful in initial parameter setting of health care providers. Springer Verlag New York Inc; 1999.
the network. The result shows that the best area under the ROC [13] Karahaliou Anna N, Boniatis Ioannis S, Skiadopoulos Spyros G, Sakellaropoulos
curve is found to be 0.96853 for real clinical Database. The Filippos N, Arikidis Nikolaos S, Likaki Eleni A, et al. Breast cancer diagnosis:
analyzing texture of tissue surrounding microcalcifications. IEEE Trans Inf
PSOWNN approach generates a sensitivity of 94.167% with a spec- Technol Biomed 2008;12(6):731–8.
ificity of 92.105%. The misclassification rate is found to be [14] Kupinski MA, Giger ML. Automated seeded lesion segmentation on digital
0.063291 which is less when compared with the classification done mammograms. IEEE Trans Med Imaging 1998;17(4):510–7.
[15] Sahiner Berkman, Chan Heang-Ping, Pretrick Nicholas, Wei Datong, Mark A,
by other classifiers. When optimized learning is introduced, there
Dorit D, et al. Classification of mass and normal breast tissue – a convolution
is an improvement in classification accuracy which is 93.671%. neural network classifier with spatial domain and texture images. IEEE Trans
Therefore, PSOWNN have a great potential to be applied in auto- Med Imaging 1996;15(5):598–609.
matic detection of abnormalities in mammograms by reducing [16] Eltonsy Nevine H, Tourassi Georgia D, Elmaghraby Adel S. A concentric
morphology model for the detection of masses in mammography. IEEE Trans
the misclassification rate. Med Imaging 2007;26(6):880–9.
The performance measures of the optimally tuned wavelet neu- [17] Zheng Baoyu, Qian Wei, Clarke Laurence P. Digital mammography: mixed
ral network classifier in comparison with the other classifiers are feature neural network with spectral entropy decision for detection of
microcalcifications. IEEE Trans Med Imaging 1996;15(5):589–97.
presented in Table 3. Summarizing the results for real clinical [18] Yu Sung-Nien, Huang Yu-Kun. Detection of microcalcifications in digital
mammogram the classification accuracy of PSOWNN is higher than mammograms using combined model-based and statistical textural features.
that of the other well known classifier models. This is because of Exp Syst Appl 2010;37(7):5461–9.
[19] Sutton RN, Hall EL. Texture measures for automatic classification of pulmonary
the fact that the PSOWNN incorporates the wavelet neural network disease. IEEE Trans Comput 1972;C-21(7):667–76.
and optimally designs the neural network using swarm intelli- [20] Wang Ted C, Karayiannis Nicolas B. Detection of microcalcifications in digital
gence algorithm. This superior performance makes PSOWNN suit- mammograms using wavelets. IEEE Trans Med Imaging 1998;17:498–509.
[21] Brem Rachel F. Clinical versus research approach to breast cancer detection
able for efficiently detecting abnormalities in mammograms. with CAD: where are we now? Am J Roentgenol 2007;188(1):234–5.
[22] Christodoulou Christodoulos I, Michaelides Silas C, Pattichis Constantinos S.
Multifeature texture analysis for the classification of clouds in satellite
6. Conclusion imagery. IEEE Trans Geosci Remote Sens 2003;41(11):2662–8.
[23] Gonzales RC, Woods. Digital image processing, 2nd ed. Upper Saddle River, NJ:
The novel approach presented in this paper demonstrated that Prentice Hall; 2002.
[24] Laws KJ. Texture energy measures. In: Proceeding DARPA image understanding
the PSOWNN classifier produces an improvement in classification workshop; 1979. p. 47–51.
accuracy to the problem of computer-aided analysis of digital [25] Kocur CM, Rogers SK, Myers LR, Burns T, Kabrisky M, Hoffmeister JW, et al.
mammograms for breast cancer detection. The algorithm devel- Using neural networks to select wavelet features for breast cancer diagnosis.
Eng Med Biol Mag IEEE 1996;15(3):95–102. 108.
oped here classifies mammograms into normal & abnormal. First, [26] Maglogiannis I, Sarimveis H, Kiranoudis CT, Chatziioannou AA, Oikonomou N,
the ROI of the image is chosen then laws texture features are ex- Aidinis V. Radial basis function neural networks classification for the
tracted and classified using PSOWNN. The optimized WNN based recognition of idiopathic pulmonary fibrosis in microscopic images. IEEE
Trans Inf Technol Biomed 2008;12(1):42–54.
classifiers using the properties of both wavelet and neural network
[27] Sun Bing-Yu, Zhu Zhi-Hua, Li Jiuyong, Linghu Bin. Combined feature selection
provide good classification accuracy by reducing the false positives and cancer prognosis using support vector machine regression. IEEE/ACM
and false negatives. The PSOWNN classifier designed using PSO Trans Comput Biol Bioinform 2011;8(6):1671–7.
algorithm applied to WNN is investigated for detecting breast can- [28] Zhang Q, Benvniste A. Wavelet networks. IEEE Trans Neural Networks
1992;3(6):889–98.
cer in mammograms. The results give better classification accuracy [29] Sadri AR, Zekr M, Sadri S, Gheissari N, Mokhtari M, Kolahdouzan F.
than the traditional classifiers. The optimized wavelet neural net- Segmentation of Dermoscopy images using wavelet networks. IEEE Trans
work accelerates the convergence of the error back propagation Biomed Eng 2013;60(4):1134–41.
[30] Zhang Jun, Walter GG, Miao Y, Nagi Wan, Lee Wayne. Wavelet neural networks
algorithm and also it avoids major disruptions in the direction of for function learning. IEEE Trans Sig Process 1995;43(6):1485–97.
learning.
52 J. Dheeba et al. / Journal of Biomedical Informatics 49 (2014) 45–52

[31] Obuchowski NA. Receiver operating characteristics curves and their use in [35] Hupse R, Karssemeijer N. Use of normal tissue context in computer aided
radiology. Radiology 2003;229(1):3–8. detection of masses in mammograms. IEEE Trans Med Imaging
[32] Kennedy J, Eberhart R. Particle swarm optimization. Proc IEEE Conf Neural 2009;28(12):2033–41.
Networks 1995;4:1942–8. [36] Youden WJ. An index for rating diagnostic tests. Cancer 1950;3(1):32–5.
[33] Zhou Xiao-Hua, McClish Donna K, Obuchowski Nancy A. Statistical methods in [37] Dheeba J, Tamil Selvi S. A swarm optimized neural network system for
diagnostic medicine. Wiley; 2002. classification of microcalcification in mammograms. J Med Syst
[34] Huo Z, Giger ML, Vyborny CJ. Computerized analysis of multiple- 2012;36(5):3051–61 [Springer].
mammographic views: potential usefulness of special view mammograms [38] Dheeba J, Tamil Selvi S. An improved decision support system for detection of
in computer-aided diagnosis. IEEE Trans Med Imaging 2001;20(12): lesions in mammograms using differential evolution optimized wavelet neural
1285–92. network. J Med Syst 2012;36(5):3223–32 [Springer].