J Med Syst (2012) 36:2901–2907
DOI 10.1007/s10916-011-9768-0
ORIGINAL PAPER
Prediction of Breast Cancer Using Artificial Neural Networks
Ismail Saritas
Received: 31 May 2011 / Accepted: 2 August 2011 / Published online: 12 August 2011
# Springer Science+Business Media, LLC 2011
Abstract In this study, an artificial neural network (ANN)
was developed to determine whether patients have breast
cancer or not. Whether patients have cancer or not and if
they have its type can be determined by using ANN and BI-
RADS evaluation and based on the age of the patient, mass
shape, mass border and mass density. Though this system
cannot diagnose cancer conclusively, it helps physicians in
deciding whether a biopsy is required by providing
information about whether the patient has breast cancer or
not. Data obtained from 800 patients who were diagnosed
with cancer definitively through biopsy. The definitive
diagnosis corresponding to each patient and the data from
ANN model results were investigated using Confusion
matrix and ROC analyses. In the test data of the ANN
model that was implemented as a result of these analyses,
disease prediction rate was 90.5% and the health ratio was
80.9%. It is seen from these high predictive values that the
ANN model is fast, reliable and without any risks and
therefore can be of great help to physicians.
Keywords Artificial neural network . BI-RADS . Breast
cancer . Breast cancer prediction
Introduction
Breast cancer is the most frequently encountered cancer type
in women and the second most terminal one after the lung
cancer [1, 2]. It is also seen in men though not so frequently.
I. Saritas (*)
Department of Electronics and Computer Education,
Technical Education Faculty, Selcuk University,
Konya, Turkey
e-mail: isaritas@selcuk.edu.tr
Cancer risk increases with age. Mammography is a method
that is used in scanning asymptomatic women and can
identify cancer breast cancer in an initial stage (a-c). The
combined use of mammography and physical examination
significantly reduces mortality in breast cancer.
Breast calcifications are often seen in mamographies and
most of them are benign calcifications, but especially
calcifications smaller than 1 mm are the most precise
mammographic finding of early breast cancer. 70% of in
situ carcinomas manifest themselves only through micro-
calcifications [3, 4]. Therefore, identification of calcifica-
tions constitutes an important field of mammographic
evaluation [5].
Microcalcifications are an important characteristic of
various breast lesions that also include carcinomas. The
appearance (size and shape), number and distribution of
microcalcifications are generally very important for radiol-
ogists to determine whether the lesion that was identified
through mammography is benign or malign. However, the
lesion needs to be examined histopathologically for a
definitive diagnosis [6].
Although interpretations of microcalcifications may exhibit
variations among radiologists, an effort must be made to make
evaluation standard as much as possible and an unambiguous
conclusion must be given to clinicians. In 1993, ‘The American
College of Radiology’ (ACR) developed a standard reporting
system called ‘Breast Imaging Reporting and Data System’
(BI-RADS) in order to improve communication between
clinicians and radiologists and provide a common terminology
among radiologists [7]. A qualified evaluation of breast
density (thickness) is possible through BI-RADS, which is
commonly used these days. A sample mammogram image of
the 4 categories of breast density is given in Fig. 1 [8].
BI-RADS was updated in 2011 and divided into 6
categories (Table 1) [9].
2902
Fig. 1 Representations of the 4 breasts imaging reporting and data system (BI-RADS) breast density qualitative and quantitative assessments
The shape, boundary and density of the tissue are
determined through ultrasound and mammography. This
helps the doctor to decide on biopsy. On the other hand, it
has been reported that about 10–30% of all breast biopsies
indicate malign breast cancers [10]. Unnecessary breast
biopsies lead to both redundant expenses for tests as well as
major psychological and physical disorders for the patient
[10, 11].
In recent years, computer assisted diagnostic (CAD)
systems have been designed based on BI_RADS [12]
standards to determine tissue deformation so that the doctor
can follow the suspected area or perform a breast biopsy.
Generally, BI_RADS features are gathered from different
radiology centers for a suspected area seen in the
mammography for different BI-RADS features such as
shape and boundary of the mass provided by differentially
trained physicians.
Initially, an ANN model was devised to draw conclusions
from BI_RADS definitions [13, 14]. Then, alternative
approaches based on Bayesian network and case-based
Table 1 BI-RADS mammography classification
BI-RADS category Definition
0 Incomplete assessment
1 Negative study
2 Benign
3 Probably benign
4 Suspicious
5 Highly suggestive of malignancy
6 Known biopsy-proven malignancy
J Med Syst (2012) 36:2901–2907
reasoning was designed (CBR) [15–19]. The advantage of
CBR is that it is a process of clear judgment that leads to
diagnosis. A CBR makes inferences on the basis of a stored
database. ANN, on the other hand, is an intelligent system
that continuously renews itself by using the database
similarly. Therefore, it yields more accurate predictive results
for diagnosis than CBR.
Besides image processing, ultrasonography and physical
examination, there is a need for performing a biopsy and
making a definitive diagnosis to diagnose breast cancer.
Despite the need for a biopsy to make a definitive
diagnosis, patients with a low risk of cancer tend to avoid
it due to the complications and high costs that it may involve.
Therefore, owing to the stated reasons, patients may choose a
different method that can make an accurate decision before
they have a biopsy performed on them. To attain this goal, an
ANN model was devised to help doctors decide whether a
biopsy is required or the patient should be monitored closely
according to the patients’ BI-RADS values, age, shape of the
mass, mass boundary and mass density.
Risk of malignancy Recommended follow-up
N/A Further workup
N/A Repeat mammogram in 1y
N/A Repeat mammogram in 1y
<2% Repeat mammogram in 6 month
20% Biopsy should be considered
90% Appropriate action should be taken
N/A Appropriate action should be taken
J Med Syst (2012) 36:2901–2907
Materials and methods
In this study, literature data provided by M. Elter, R.
Schulz-Wendtland and T. Wittenberg were used [10, 20].
Those of the data that involved ambiguity were removed
and thus 822 out of 956 data were used.
MATLAB Neural Network Toolbox software was used to
develop the ANN model and make the analyses.
Artificial neural network (ANN)
Artificial Neural Networks are computer systems that have
been developed for the purpose of automatically performing
skills such as generating new information through learning
and forming and discovering new information, which are
characteristics of the brain.
Generally, it consists of three layers, namely an input
layer, one or more hidden layers and an output layer
(Fig. 2). Each layer has a certain number of elements that
are bound to each other called neurons or knots. Each
neuron is connected to the other via link weights and
accompanying communication links. Signals travel along
the neurons on link weights. Each neuron receives multiple
inputs from other neurons in proportion to their link
weights and generates an output signal that can also be
generated by other neurons [21, 22].
The network is subjected to two processes, namely training
and test, to develop an ANN model. In the case of training, the
network is educated for an output prediction depending on the
input data. In the case of test, on the other hand, the network is
Fig. 2 Structure of an ANN
2903
tested for stopping or for saving training data and is used for
predicting an output.
The process of network training is stopped when the
error that is being tested has reached the desired tolerance
value [21, 22].
Back-Propagation (BP) algorithm is the most popular
algorithm which also has the largest area of use. BP
consists of two phases, i.e. feedforward and back-
propagation procedures.
During feedforward, information that is subjected to
processing from the input layer to the output layer is
generated. In the case of back-propagation, on the other
hand, the difference between the network output value
obtained from the feedforward procedure and the desired
value is compared with the desired difference tolerance and
the error in the output layer is calculated. The obtained
error is back-propagated to update the links in the input
layer neurons [21, 23].
BP training algorithm is a ramp descent algorithm. BP
algorithm is used to improve the performance of the
network by reducing the total error through changing the
weights along the ramp. Training is stopped when the mean
square error (MSE) values stop decreasing and when there
is an increase in these values, which is an indication of
over-training [21, 24].
1X n
MSE% ¼ ðdi  Oi Þ2 ð1Þ
n i¼1
Here, di is the target or true value, Oi is the network output
value or predicted value, and n is the output data number.
2904
Application of artificial neural network to the data
The purpose of this devised ANN is to predict severity
(benign or malign) of a mammographic mass tissue by
using a patient’s age and BI-RADS features.
The data set that was used involved BI-RADS values,
the patient’s age and three BI-RADS features and area of
significance. These data were defined on full range digital
mammogram for 399 benign and 423 malign tissues in the
radiology institute of Erlangen-Nuremberg University in
2000–2006 [10, 20]. 169 of the data were selected
randomly to test the model later. The remaining 653 data,
on the other hand, were selected randomly to be used in the
training of the model, i.e. 80% for training, 5% for
validation and 15% for testing.
Each sample was examined by the physician and defined
as follows [10, 20].
1. BI-RADS evaluation: 1–5 (consecutive, not a prediction!)
2. Age: Age of the patient (digital)
3. Figure 1: mass shape: round = 1 oval = 2 round
protruding = 3 irregular = 4 (nominal)
4. Mass edge boundaries: boundaries are defined = 1,
microlobulated = 2, obscured = 3, ill-defined = 4,
speculated = 5 (nominal)
5. Density: Mass density high = 1, iso = 2, low = 3, fat-
containing = 4 (ordinal)
6. Nature: benign = 0 malign = 1 (binominal, goal field!)
In this study, a feedforward network structure containing
an entry layer, a hidden layer and output layer (Fig. 6) was
used. After the ANN structure was developed, it was
normalized within the 0–1 value set using Eq. 2 [21] in
order to improve the training characteristics of the data set.
x  xmin
xnorm ¼ ð2Þ
xmax  xmin
The training data set was used to determine ANN neuron
and bias weight values. Training was repeated by changing
the number of neutrons in the hidden layer and the iteration
Table 2 Parameters that
were used in ANN Parameters
The number of neurons in the input layer
The number of hidden layers
The number of neurons in the hidden layer
The number of neurons in the output layer
Learning rate (α)
Training rate coefficient (β)
Learning algorithm
Transfer function
J Med Syst (2012) 36:2901–2907
number in order to obtain the lowest error value. After the
most suitable network structure was determined, the trained
algorithm was applied to the test data set. The ANN
parameters that were used are given in Table 2.
The ANN was created by using the BP algorithm which
has minimum error delivery and 9 neurons in its hidden
layer. A mathematical formula was obtained using ANN
and Eq. 6.
For Eq. 4, NETj is obtained using Eq. 3.
 
ðWi Þi;1  BIRADS þ ðWi Þi;2  Age þ ðWi Þi;3  Shape þ
NETj ¼
ðWi Þi;4  Margin þ ðWi Þi;5  Density  2:1191
ð3Þ
The constants which are link weight values (W1)i,j for the
BP algorithm, which has 9 neurons in its hidden layer, are
given in Table 3. In the formula given above, NETj is the
total of the multiplication of input parameters and their
weights. i and j, which are subscripts, are input and hidden
neuron numbers respectively.
In the ANN input layer, as BI-RADS features and age
entry parameters, 9 neurons were used in the hidden layer
and the Severity diagnosis was used in the output layer. 9
equation pieces were used for NET1-NET9 and F1-F9, as
total and activation functions respectively.
Here, the transfer function that was used for this
approach is the TANSIG transfer function that is given in
Eq. 4.
2
Fj ¼ 1 ð4Þ
1 þ eð2NETj Þ
For Eq. 6, A is obtained using Eq. 5.
0 1
0:4421  F1  0:3955  F2  0:4658  F3 þ 0:3157  F4
A ¼ 1:1796  F5  0:0061  F6  0:2972  F7 þ 0:4249  F8 A
@
0:2708  F9  2:1191
ð5Þ
Properties
5
1
9
1
0.3
0.3
Gradient reduction algorithm (trainlm)
Logarithmic sigmoid (tansig)
J Med Syst (2012) 36:2901–2907 2905

Table 3 Weight values for Eq. 7

Number of neurons BI-RADS W1(i,j) Age W1(i,j) Shape W1(i,j) Margin W1(i,j) Density W1(i,j)
in the hidden layer

1 1,1627 −1,2199 −1,0579 −0,4368 −0,8746

2 −0,3600 0,2081 1,1471 0,9569 1,3134
3 −0,9299 −1,2773 0,4513 −0,0430 1,2915
4 1,2861 −1,1498 −0,9214 −0,4671 −0,7741
5 −1,9271 −0,0753 −0,2703 −0,5172 −0,8155
6 1,0994 −0,9916 −1,4366 0,0361 0,6090
7 −0,3945 −1,2386 −1,1503 −1,1803 0,3392
8 1,2736 1,1961 0,1942 0,9558 −0,7063
9 −0,1239 0,9133 −1,3526 1,0869 −0,8882

Equation 6 is severity result. Predictive value of the positive result (PVP) is the actual
possibility of being sick when the diagnostic test passed the
2 judgment of sick (Eq. 10).
Severity ¼ 1 ð6Þ
1 þ eð2AÞ
A
PVP ¼ ð10Þ
AþB

The predictive value of the negative result (PVN) is the

Results and suggestions
possibility of being healthy when the diagnostic test said
healthy (Eq. 11).
Confusion [25] matrix and ROC [24] analyses were
performed in order to see the success of the study that D
was conducted. PVN ¼ ð11Þ
DþC

Of the 693 training data, the ANN that was performed

Confusion matrix analysis classified 285 out of 344 malign data and 296 out of 349
benign data to be successful (Table 4).
In the Confusion matrix analysis: When the diagnostic test that was performed on the
Sensitivity is the ability expressed with the rate at which training data in Table 3 was examined, the following results
the system can determine real patients among those that were reached:
were definitely diagnosed to be sick (Eq. 7). Using Eq. 7, 8, 9, 10 and 11 successively, these values
were found to be: sensitivity=0.843, specificity=0.834,
A accuracy=0.838 and PVP=0.828 and PVN=0.848.
Sensitivity ¼ ð7Þ Moreover, of the 131 test data that were not subjected to
AþC
ANN training, ANN classified 55 out of 68 malign data and
Specificity is the ability expressed with the rate at which 57 out of 63 benign data to be successful (Table 5).
the system can determine real healthy among those that When the diagnostic test that was performed on the
were definitely diagnosed to be healthy (Eq. 8). training data in Table 5 was examined, the following results
were reached:
D
Specificity ¼ ð8Þ
DþB
Table 4 Training data set confusion matrix
Accuracy is the actual total correct diagnoses as sick and Classification Malign Benign Sum of rows
healthy (Eq. 9)
Malign 285 (A) 59 (B) 344
Benign 53 (C) 296 (D) 349
AþD
Accuracy ¼ ð9Þ Sum of Columns 338 355 693
AþBþCþD
2906 J Med Syst (2012) 36:2901–2907

Table 5 Test data set confusion matrix

Classification Malign Benign Sum of rows

Malign 55 (A) 13 (B) 68

Benign 6 (C) 57 (D) 63
Sum of Columns 61 70 131

Using Eq. 7, 8, 9, 10 and 11 respectively, the following

values were found: sensitivity=0.902, specificity=0.814,
accuracy=0.855, PVP=0.809 and PVN=0.905.

ROC analysis

The basic idea behind medical tests is to estimate the patient’s

Fig. 3 Graph of the ROC analysis of training data
possibility of being sick on the basis of their test results. ROC
Analysis is used to determine the true accuracy of medical
diagnostic tests. Of the terms used in this analysis, sensitivity whereas the area below ANN test data YSA (Fig. 4) ROC
is the ratio of the sick persons (True Positive) whose tests were curve was determined to be 0.870017. The fact that these
positive (sick) to the number of patients whereas specificity is areas are large (close 1) indicates that the study that was
the ratio of healthy persons (False Positive) whose tests were performed was successful in breast cancer diagnostic
negative (healthy) to the number of healthy persons [21, 22, predictions.
25]. ROC analysis is a standard approach which is used to According to the both analyses, it is observed that the
determine the sensitivity and specificity of the diagnostic BI-RADS features that were used as input parameters and
procedure. To this end, ROC curves that will define the age data can be used in making accurate predictions in
relationship between the sensitivity and specificity of the breast cancer severity diagnosis.
diagnosis are used. The axes of the curves are TP (calling the When the study that was performed is compared with the
sick sick) and FP (calling the healthy sick). The curves are studies in the relevant literature that were conducted using
between the limits of 0 and 1, and while proximity to different methods, it is seen that it can be used as a
coordinate y and upper boundary indicates a successful test, supplementary tool that may eliminate unnecessary biopsy.
the curves that have an inclination of 45o indicate an Moreover, better breast cancer prediction results can be
unsuccessful test [21, 22, 25] obtained by increasing the number of patient data, adding
Thus, the success of the test can be determined by appropriate input parameters and using methods of artificial
examining ROC curves. In a successful test, the area below intelligence that can be used in conjunction with ANN.
the curves is expected to be greater. Figure 3 shows the
ROC curves that indicate accuracy of the data set used in
this study at the stage of training while Fig. 4 shows ROC
curves that indicate test accuracy.
Using the equation obtained in the study (Eq. 6),
patients’ BI-RADS features and age data and ANN severity
prediction value can be determined quickly and easily.
According to the test data set confusion matrix analysis, it is
observed that the ratio of distinguishing the sick from the true
sick is 90.2%, the ability to distinguish the healthy from the
true healthy is 81.4% and the ratio of total number of correct
diagnoses of the test as sick and healthy is 85.5%. According to
these results, when the ANN predictive value passed the
judgment of sick, the possibility of being true sick is 80.9%.
Also, when the ANN predictive value passed the judgment of
healthy, the possibility of being true healthy is 90.5%.
According to the ROC analysis, the area below ANN
training data (Fig. 3) ROC curve was found to be 0.850017, Fig. 4 Graph of the ROC analysis of test data
J Med Syst (2012) 36:2901–2907
Acknowledgements This study was supported by Selcuk University
Coordination Office of Scientific Research Projects (BAP). Moreover,
I would like to express my heartfelt thanks for Prof. Dr. Unal SERT of
Selcuk University, Meram Medicine Faculty, who was my advisor and
helped me in the evaluation of breast cancer data.
References
1. Firat, D., and Hayran, M., Cancer Statistics in Turkey and in the
World (1990–1992). Turkish Association for Cancer Research and
Control. İz Press, Ankara, 1995.
2. Şengelen, M., Kutluk, T., and Fırat, D., Cancer statistics in Turkey
and in the World (1996–2003). Turkish Association for Cancer
Research and Control. İz Press, Ankara, 2007.
3. Stomper PC, Connolly JC, Meyer JE, et al. (1989) Clinically
occult ductal carcinoma in-situ detected with mammography:
Analysis of 100 cases with radiologic-pathologic correlation.
Radiology 172–235
4. Holland, R., Ductal carcinoma in situ (DCIS). Eur. Radiol.
10:327–330, 2000.
5. American College of Radiology, Breast imaging reporting and
data system (BI-RADS), 3rd edition. American College of the
Radiology, Reston, VA, 1998.
6. Erkul, Z. K., Erkuş, M., Taşkın, F., and Meteoğlu, İ., Liesegang
ring calcification in breast biopsy: case report. J. Breast Healthy. 1
(1):22–24, 2005.
7. Gülsün, M., Demirkazık, F. B., Köksal, A., and Arıyürek, M.,
According to BI-RADS assessment of breast microcalcifications
and to investigate the agreement between reviewers. Off. J.
Turkish Soc. Radiol. 8(3):358–363, 2002.
8. Renee, W., Pinsky, M. D., Mark, A., and Helvie, M. D., Medscape
mammographic breast density: Effect on imaging and breast
cancer risk: Breast density measurement. JNCCN-J. Natl. Compr.
Cancer Netw. 8:1157–1165, 2010.
9. Edward, T. B., Rick, K., and Robert, R., Conn’s current therapy.
Elsevier INC, Saunders, Philadelphia, 2011.
10. Elter, M., Schulz-Wendtland, R., and Wittenberg, T., The prediction
of breast cancer biopsy outcomes using two CAD approaches that
both emphasize an intelligible decision process. Med. Phys. 34
(11):4164–4172, 2007.
11. Nassif, H., Page, D., Ayvaci, M., Shavlik, J., and Burnside, E. S.,
Uncovering age-specific invasive and DCIS breast cancer rules using
inductive logic programming. In: Veinot, T. (Ed.), Proceedings of the
2907
1st ACM International Health Informatics Symposium (IHI ‘10).
ACM, New York, pp. 76–82, 2010.
12. American College of Radiology, Breast imaging reporting and
data system (BI-RADS), 4th edition. American College of
Radiology, Reston, VA, 2003.
13. Baker, J. A., Kornguth, P. J., Lo, J. Y., Williford, M. E., and Floyd, C.
E., Breast cancer: Prediction with artificial neural network based on
BI-RADS standardized lexicon. Radiology 196:817–822, 1995.
14. Markey, M. K., Lo, J. Y., Vargas-Voracek, R., Tourassi, G. D., and
Floyd, C. E., Perception error surface analysis: A case study in
breast cancer diagnosis. Comput. Biol. Med. 32:99–109, 2002.
15. Floyd, C. E., Lo, J. Y., and Tourassi, G. D., Case-based reasoning
computer algorithm that uses mammographic findings for breast
biopsy decisions, AJR. Am. J. Roentgenol. 175:1347–1352, 2000.
16. Bilska-Wolak, A. O., and Floyd, C. E., Investigating different
similarity measures for a case-based reasoning classifier to predict
breast cancer. Proc SPIE 4322:1862–1866, 2001.
17. Bilska-Wolak, A. O., and Floyd, C. E., Development and evaluation
of a case-based reasoning classifier for prediction of breast biopsy
outcome with BI-RADS lexicon. Med. Phys. 29:2090–2100, 2002.
18. Bilska-Wolak, A. O., Floyd, C. E., Lo, J. Y., and Baker, J. A.,
Computer aid for decision to biopsy breast masses on mammography:
Validation on new cases. Acad. Radiol. 12:671–680, 2005.
19. Markey, M. K., Fischer, E. A., and Lo, J. Y. Bayesian networks of
BIRADS descriptors for breast lesion classifications. International
Conference of the IEEE Engineering in Medicine and Biology
Society. San Francisco, California 3031–3034, 2004
20. UCI Machine Learning Repository: Data Sets 2007. http://archive.ics.
uci.edu/ml/datasets/Mammographic+Mass, Accessed 29 March 2011
21. Saritas, I., Ozkan, I. A., and Sert, I. U., Prognosis of prostate
cancer by artificial neural networks. Expert Syst. Appl. 37
(9):6646–6650, 2010.
22. Ronco, A. L., and Fernandez, R., Improving ultrasonographic
diagnosis of prostate cancer with neural networks. Ultrasound
Med. Biol. 25(5):729–733, 1999.
23. Allahverdi, N., Expert systems. An artificial ıntelligence application.
Atlas Press, Istanbul, 2002.
24. Türker, N., Tokan, F., and Yıldırım, T. Determination of artificial
neural networks performances by roc analysis on the diagnosis of
heart disease. Biomedical Engineering National Congres, BİYO-
MUT 2005, İstanbul 206–208, 2005.
25. Wichard, J.D., Cammann, H., Stephan, C., and Tolxdorff, T.
Classification models for early detection of prostate cancer.
Hindawi Publishing Corporation Journal of Biomedicine and
Biotechnology 7, 2008.