Professional Documents
Culture Documents
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 93
Abstract- Breast cancer occurs mostly in women and figures states that almost one out of eight women suffers from this disease. Breast
cancer is second only to lung cancer that causes the majority of deaths among women and is difficult to identify accurately and treat.
Diagnostic computational tools can be used effectively, with high degree of accuracy, which may help in improving the specificity and
sensitivity of diagnoses. There are two known types of breast cancer namely benign and malignant which are cured mostly by mammography,
FNA (Fine Needle Aspirate) and surgical biopsy. In this research work, our main objective is to develop and compare a diagnostic system for
diagnosis; prognosis and prediction of breast cancer with the usage of the computational intelligence namely Back Propagation (BPA),
Learning vector Quantization (LVQ), Support vector machine (SVM) and Adaptive Neuro-fuzzy Inference System (ANFIS). These modeling
tools can highlight the important features that play pivotal roles in the classification and aid physicians to diagnose and prognosticate breast
cancer. The database used in developing the above discussed model in this research work is adopted from the University of Wisconsin (UCI)
Machine Learning Repository. Experimental result shows that different models give optimal performance for the data set. However, all the
models are able to solve the problem up to a greater extent but a maximum accuracy of 96.50% is being achieved on the usage of LVQ which
outdo all the other models.
Keywords- Artificial Neural Networks, Breast Cancer, Expert System, Back Propagation Algorithm, Learning vector Quantization, Support
vector machine and Neuro-fuzzy Inference System.
of care as well as patient and public opinion
1. Introduction satisfaction. Recently, computer‐based medical data
Breast cancer has become one of the most perilous processing research has yielded methods and tools for
carcinomas in the world among middle‐aged and old managing the task away from the hospital management
age women. Breast cancer is the second leading cause of level and closer to the desired disease and patient
cancer deaths in women worldwide and occurs in management level.
nearly one out of eight women [1]. Breast cancer occurs
mostly in women but hardly ever it is found to occur in 2. Literature Review
men. Computational intelligence is a stimulating field
and from the last decade, Artificial Neural Networks Yuanjiao et al. proposed a technology to extract micro‐
(ANN) has become important in medical applications calcifications clusters with accurate edge effects which
due to their accuracy for predictive inference, with canʹt be detected by the naked eye on mammograms to
potential to support large data sets. Feedforward obtain much more hidden information in order to help
Neural Network (FNN) is a kind of ANN, which has a the doctors in diagnosing breast cancer [2]. Another
better structure but due to slow training rate, easy to important research was been carried out by Heng‐Da et
trap local minimum point and poor ability in global al. who has proposed a computerized micro‐
search prevents its usage as alone. Learning involves calcification detection based on fuzzy logic, vibro‐
the extraction of rules or patterns from the set of data. acoustography and probabilistic neural network on
Further the time and memory requirement would be mammograms for breast cancer [3]. Image feature
reasonably less as the system has already summarized extraction was utilized to retrospectively analyze
the data into some patterns or rules. In the computer‐ screening mammograms taken prior to the detection of
aided decision system, a system is considered as a malignant mass was done by Mohd. Sameti et al.[4].
effective and accurate if it shows a very high Al Mutazet al. uses statistical texture features for breast
generalizing capability which would help a physician cancer detection and few ANN [5].
to diagnose a patient correctly. However, unfortunately T.Z. Tan et al.uses thermogram which is a promising
clinicians can act in different ways depending on their front‐line screening tool to warn women of cancer [6].
knowledge and experience which highlighted the need Laufer et al. proposed a modified self‐organizing map
to introduce diagnostic tools to support the scientific with nonlinear weight adjustments to reduce number of
homogeneity and accountability of healthcare decisions unnecessary biopsies with a minimal number of
and actions. The benefits expected from such actions subsequent [7]. Taio et al. used Kohonen self organizing
include an overall reduction in cost, improved quality map and multilayer perceptron trained with the
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 3, MARCH 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 94
backpropagation algorithm [8]. Xiong et al. used Collect data set for Breast Cancer
statistics methods like PCA, PLS linear regression
analysis, data mining methods, and hybrid system of
Data Feature Extraction and analysis.
combination rough set and probabilistic neural Processing Removal OR Filling of missing data.
network. Probabilistic neural network to perform Data Normalization.
supervised classification and rough sets was able to
reduce the number of attributes in the dataset without Models
Simulated
sacrificing classification accuracy in [9, 10, 11].
Karabatak et al. used ANN and ANFIS with linear
Training and Testing Phase
discriminant analysis and principal component analysis
for diagnosis of breast cancer. [12, 13]. Ky Van Ha et al.
Diagnosis Decision
constructed a hierarchical evolutionary RBF network
and employed it to detect the breast cancer. [15, 16].
The optimum network for classification of breast cancer Benign Malignant
cells was found using Hybrid Multilayer Perceptron
(HMLP) network. [17, 18].
Performance Comparison of computational models
Jain et al. used fuzzy‐logic, Hybrid Neuro‐Fuzzy
generator based on the Knowledge Oriented Design
(KOD) concept and Cooperative Neuro‐Fuzzy systems Figure 1: Block Diagram of Complete Methodology
using Genetic Algorithms were used for the
classification (diagnosis) of breast cancer [19, 20]. Pena‐ 3.1. Diagnosis using Back Propagation Algorithm
Reyes et al. proposed fuzzy‐genetic breast cancer
identification. [21]. Seker et al. proposed a methodology In back propagation, the partial derivatives of the cost
with neural network, fuzzy logic, FK‐NN and statistical function with respect to the free parameters of the
method to prognostic analysis of cytometric image network is being determined by back propagation error
data. [22]. signals through the network, layer by layer. The feed‐
forward neural network architecture used in this
experiment consists of a single hidden layer along with
3. Methodology
input and output. The employed neural network is a
In this research work, computational intelligence feed forward neural network and it is shown in the
models have been used for the diagnosis and prediction following figure
of breast cancer. The learning takes place through an
iterative process of weight adjustments applied to its
initial weight after epoch iteration of the learning
process. Figure 1, shows the overall diagram of
complete methodology. Mammography, biopsy and
FNA database of breast cancer microscopic and clinical
tests reports are used in our diagnostic system which
helps in prediction of cancer as either Benign or
Malignant. In this research work, we have employed
ANN like Back Propagation (BPA), Support vector
machine (SVM), learning vector Quantization (LVQ)
Figure 2. Back Propagation Multi‐layer neural network
and Adaptive Nero‐fuzzy Inference System (ANFIS) for
structure.
training of the dataset. Among all the above mentioned
ANN, LVQ achieved the 96.50% of accuracy which
The transfer function in hidden layer neurons and
comes out to be highest.
output layer neurons are sigmoid and purelin
respectively. The performance function used was Mean
Sum‐squared Error (MSE).
For hidden units, we must propagate the error back
from the output nodes; equation (1) gives us the
equation of error back propagation
E neti y j
j
(1)
iPj neti y j net j
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 3, MARCH 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 95
A1 A1 R1
x1 R1
x1
x1
A2 R2
x1 A2 R2
wR3 C1
C1
Figure 3. Learning Vector Quantization Neural A3 A3 R3 R3 wR6
y
Network Structure wR1
B1 B1 R4 R4 wR2
x2 C2
Here, input layer consists of data provided. In the wR4
C2
R5
competitive layer, each competitive unit corresponds to x2 x2 wR5
B2 B2 R5
a cluster and the output layer has the neurons equal to x2
R6
the no of classes. B3
B3 R6
According to LVQ algorithm,
(a) If w xi
c
wc ( n 1) wc ( n) n [ xi wc ( n)]
Figure 4. Adaptive Neuro‐fuzzy Inference Systems
where, 0 < n < 1. (2)
else w xi The ANFIS can be trained by a hybrid learning
c
algorithm [28]. In the forward pass the algorithm uses
wc ( n 1) wc ( n) n [ xi wc ( n)] (3) least‐ squares method to identify the consequent
Where, {w j}lj1 represent the set of Voronoi vectors, parameters while in the backward pass the errors are
propagated backward and the premise parameters are
w c
represent the class associated with the Voronoi updated by gradient descent [27].
vector, {x i }iN1 represent the set of input vectors, Layer 1 is the input layer.
x represent the class label of the input vector xi, and
i yi(1) xi(1) (4)
n denotes the learning constant with the number of
iterations n.
Layer 2 is the fuzzification layer.
(b) The other Voronoi vectors are not modified. b
0, if xi( 2 ) a
LVQ network has an advantage over perceptron as they 2
can classify any set of input vectors whereas perceptron
2 xi a
( 2)
( 2) b b
can classify only linearly separable sets of input vectors. yi 1 , if a xi( 2 ) a
The only requirement is that the competitive layer must
b 2 2 (5)
have enough neurons and each class must be assigned b
enough competitive neurons [23][24][25]. 0, if xi( 2 ) a
2
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 3, MARCH 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 96
subject to (for any 1=1,2,3,………….,n)
Layer 3 is the inference layer. (14)
yi(3) x1(i3) x2(3i ) xki(3) (6) 4. Experiments and ( Results
The results are 6 measured against the following
diagnostic performance measures:
)
Layer 4 is the output membership layer 1. True positive (TP): The number of positive
cases correctly detected,
y
i
( 4)
x ( 4)
1i x ( 4)
2i x ( 4)
li (7)
2. True negative (TN): The number of negative
cases correctly detected,
3. False positive (FP): The number of negative
Layer 5 is the defuzzification layer. For, the datasets C1
cases diagnosed as positive, and
and C2, the weighted average would be calculated as,
4. False negative (FN): The number of positive
cases diagnosed as negative.
μ a b μ a b
y C1 C1 C1 C2 C2 C2 (8) The Wisconsin breast cancer diagnosis (WBCD)
(
μ b μ b database is the result of the efforts made at the
C1 C1 C2 C2 8
University of Wisconsin Hospital for accurately
)
diagnosing breast masses based solely on an FNA test.
The database comprises diagnosis results of 699
3.4. Diagnosis using Support Vector Machine
patients beyond which 458 are benign and 241 are
(SVM)
Support Vector Machine (SVM) is based on the linear malignant [25]. Different neural network models were
classifiers and on a general method known as “kernel simulated for different parameter settings. The various
tricks”, which is based on the concepts of duality and performance measures are summarized and
kernels that allow algorithms to be formulated such experimental results of breast cancer system using
that they could be systematically extended to non‐ computational intelligence models are shown in table 1
linear cases. and table 2 respectively. LVQ has emerged as the
Support Vector Machine (SVM) models are similar to optimal network, which uses 10 hidden neurons and
multilayer perceptron neural networks. Using a kernel 0.01 learning rate.
function these are an alternative training method for
polynomial, radial basis function and multi‐layer Table 1. Diagnostic performance measurement for
perceptron classifiers. These classifiers are found to be Breast cancer.
helpful in solving a quadratic programming problem
Cancer Total
with linear constraints as compared to others. Present Absent
Test
In SVM, let us assume the training data to be D, a set of
True False [TP +FP]
n points of the form
Positive Positive Positive
(9)
[TP] [FP]
where, ci is either 1 or −1, indicating the class to which
False True [FN+TN]
the point Xi belongs. Each Xi is a p‐dimensional real
Negative Negative negative
vector. Any hyperplane can be written as the set of
[FN] [TN]
points X satisfying
(10) [TP + FN+ TN
Total [TP+ FN] [TN + FP]
We want to choose the W and b to maximize the + FP]
margin. These hyperplanes can be described by the
equation
(11) Sensitivity (Sens.)
TP / (TP + FN)
And Specification (Spec.)
TN / (TN + FP)
(12) Accuracy (Acc.)
(TP + TN) / (TP + TN + FP +
If the training data are linearly separable then we can FN)
select the two hyperplanes of the margin in such a False Positive Rate FP / (TP+FP)
manner that there are no points between them and then (FPR)
try to maximize their distance False Negative Rate FN / (FN+TN)
for all 1≤ i ≤ n (13) (FNR)
We can put this together to get the optimization
problem:
Minimize (in w,b)
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 3, MARCH 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 97