Content uploaded by Mohsin Khan
Author content
All content in this area was uploaded by Mohsin Khan on Sep 03, 2021
Content may be subject to copyright.
2019 13th International Conference on Mathematics, Actuarial Science, Computer Science and Statistics (MACS)
Breast Cancer Classification and Proof of Key
Artificial Neural Network Terminologies
Nisar Ali, Shahab Ansari, Zahid Halim, Raja Hashim Ali, Muhammad Faizan Khan, Mohsin Khan
Faculty of Computer Science and Engineering
Ghulam Ishaq Khan Institute of Engineering Sciences and Technology
{nisarali4465, ansarishahab111, halimzahid, raja.hashim.ali1, faizan4465.khan, mykhan911}@gmail.com
Abstract—Classification is one of the interesting areas in the
academic field of Neural Networks. Artificial Neural Networks
(ANNs) have been extensively used in pattern recognition
and classification of data in the supervised and unsupervised
environment. The ANNs use advanced concepts of computer
science where a machine mimics human intelligence while
learning from possible experience. To make a machine self-
adaptive and autonomous, the machine is properly trained on a
training data-set and then subsequently tested on new data. The
excellent quality of training of ANNs typically depends on the
underlying architecture of the network they employ, for a specific
instance, a considerable number of deep layers, number of key
nodes in each distinct layer, epoch size, and activation function.
In this academic paper, the practical importance of these
architectural components is carefully investigated. This paper is
precisely about providing a solution that how ANNs can help us
in Breast Cancer Classification. Furthermore, sufficient proofs
of some extremely important terminologies used in ANNs are
also discussed which will clarify the important concepts of ANNs.
Index Terms—Artificial Neural Network, Key Terminologies,
Breast Cancer Classification, and Proofs
I. INTRODUCTION
Breast cancer is a type of disease where malignant (cancer)
cells form in the tissues of the breast. It starts when breast cells
begin to grow out of control. A tumor appears due to these
cells which can easily be seen on an x-ray. Breast cancer exists
almost entirely in women, but there is a chance that men can
get breast cancer, too [1].
Among women, the most commonly diagnosed cancer is
of the breast [2]. In the United States, one in eight women
is diagnosed with this disease in her lifetime. This disease
remains the second leading cause of cancer deaths in women.
Each year over 252,500 women in the United States are
diagnosed with breast cancer and more than 40,000 are died
because of this disease. This disease is rare in men, but it is
estimated that over 2,500 men are diagnosed with this disease
and 500 die every year [3]. On average, every two minutes a
woman is diagnosed with this fatal disease and a woman has
died every 14 minutes. Over 3.3 million survivors are alive in
the United States with breast cancer [4].
A lot of work has been done to stop this disease from
causing harm [5] [6] [7] [8]. It takes time to diagnose breast
cancer due to which it spreads and the chances of death
increase. It is very important to diagnose such a fatal disease
quickly and for this, there should be a computerized system
that will help to diagnose breast cancer faster [9]. For this
many techniques have been proposed and one of this is
Artificial Neural Networks. This method has been used widely
to classify breast cancer to help the community.
The rest of the paper is structured as follows. In section II
we introduce the Artificial Neural Network. In section III
related work is discussed. Our methodology is proposed in
section IV. Results and proofs are reported in section V.
Finally, conclusions are drawn in section VI.
II. ART IFI CI AL NEURAL NE TWORK
Artificial neural network has provided an exciting alternative
method for solving a variety of problems in different fields
of science and engineering. An artificial neural network is
defined as the interconnection pattern between the different
layers of neurons. Each connection between neurons has a
weight [8]. Each neuron is represented by a node and these
nodes are connected through edges which makes the output
of one neuron, the input of others. Two main types of neural
networks are given below:
•Feed-Forward Network
•Feedback network
A. Feed-Forward Network
Feed-forward network [10] consists of a perceptron which
are organized in the form of layers. A feed-forward network
is a simple type of neural network which consists of an input
layer, an output layer, and one or more hidden layers. Each
perceptron in a single layer has no connection with each
other. Furthermore, each perceptron in a layer is connected
to all perceptrons on the next layer. The main advantage of
this network is that it learns to evaluate and recognize input
patterns [11]. In this network, the information moves only in
one direction i.e. forward, from the input nodes data goes
through hidden nodes (if any) and then to the output nodes.
Fig 1 shows the basic model of a feed-forward network.
B. Feedback Network
In this type of network [12], the output goes back into the
network to achieve the best result. Feedback networks are dy-
namic. Signals are traveling in both directions by introducing
loops in this network. The network remains at the equilibrium
point until the input changes. Feedback networks are used by
the internal systems for the correctness of errors [13]. Fig 2
shows the basic model of the feedback network.
978-1-7281-4956-1/19/$31.00 c
2019 IEEE
C. Artificial Neural Network Training
An artificial neural network is a computing system inspired
by the human brain. A human brain learns from its surround-
ings and make decisions using knowledge. An artificial neural
network can be developed just like a human brain by training
it with proper data-set so that it can be able to solve unseen
problems related to the specific data-set. A neural network
can be trained by different methods. A feed-forward network
can be trained by using Conjugate Gradient [14] and Back
Propagation (BP) method [15], the most common learning
algorithms. On the other hand, the feedback network can be
trained by using real-time recurrent learning algorithms [16]
[17].
Fig. 1. Basic Model of Feed Forward Network
Fig. 2. Basic Model of Feedback Network
III. REL ATED W OR K
In the last decade, a lot of work has been done for the
detection of breast cancer. Today, several screening techniques
are used to detect breast cancer like positron emission tomog-
raphy (PET) [18], CT scan [19], X-Ray [20], ultrasound [21],
mammogram [22] etc. These techniques have their advantages.
Mammogram is the most reliable and popular technique. But,
this technique has some serious limitations. About 30 of
the breast grazes couldnt be spotted in mammograms during
screening. This thing led the researchers towards developing
an automated computational system for breast cancer diagnosis
[23].
Artificial neural networks [24] are used for the diagnosis
of breast cancer since the past few years and their accuracy
rate is very significant as compared to other non-computational
diagnostics. For classification Support Vector Machine (SVM)
[25], Self-Organizing Map (SOM) [26], Probabilistic Neural
Network [27], General Regression Neural Network [28], Ra-
dial Basis Network (RBN) [29] and Multi-Layer Perception
(MLP) [5] have also been applied on the said task. The
obtained results show that General Regression Neural Network
has been the most accurate rate in identifying the nature
of the input as compared to others. Its accuracy rate was
98.8% respectively [30]. Xin Yao carried out a negative
correlation algorithm [31] which was able to automatically
decompose and solve the problem. He prominently mentioned
two unique approaches like an evolutionary approach that was
able to automatically design a compact neural network. The
second unique approach was the ensemble which was precisely
in progress but was carefully designed to properly address
massive problems [32].
IV. METHODOLOGY
Wisconsin data set [33] is used in the training and testing
of Neural Networks. In this data set first column accurately
indicates the id of each row, the next nine columns indicate
authentic data and the last column accurately represents the
desired output against each row. Each specific column among
the nine columns uniquely represents a specific property of a
cell i-e Clum Thickness, Bare Nuclei, Normal Nucleoli, etc.
The output column indicates whether it is malignant, means
4 or benign, means 2. We need separating ids, authentic data,
which represent properties of a cell and output data from given
data set before we train Neural Network. There are precisely
some unknown values in the given data set which are gently
replaced by 5 in the following experiments.
Significant key parameters which are used to set architecture
in training of Neural Networks are:
•Activation Functions
•Hidden Layers
•Learning Functions
•Mean Square Error
•Epochs
V. RE SU LTS AND ANA LYSIS
The desired results of each experiment are analyzed by
checking the performance of the neural networks through
accuracy, precision, and efficiency along with proofs.
A. First Hypothesis
Training of Neural Network on half 50% of provided
data, in which 32.5% of selected data indicates Benign and
17.5% shows Malignant, would be much more effective as
compare favorably to randomly chosen data from given data
set. Successful experiments which have been precisely carry
out are adequately explained in Table I. The key parameters
which are constant for the experiment are MSE =0.01 and
Epochs =100
An immense difference in the performance of Neural Net-
works, if we compare accuracy favorably of experiments
being done under given hypothesis and randomly chosen data.
Nearly all of the successful experiments gave more than
96% of remarkable precision. The appropriate environmental
setup for the training of Neural Network, for the above-given
hypothesis, has been communicated below.
•Hidden Layers: According to the accuracy rate of the
above experiments, it has been carefully observed that
the specific number of deep layers may typically vary
between 1 and 3 and the considerable size of deep layers
can be gently set between 5 and 20. More than 3 hidden
layers would be of no use because it can be perceived
precisely that the performance decreases, a little bit or
stays constant, as we set the number of deep layers more
than 3.
•Activation Functions: logsig and tansig activation func-
tions have been properly used for hidden layers in above
experiments and tansig,purelin and softmax have been
used carefully for the output layer. The finest activation
function for hidden layer remained tansig while for output
layer purelin and tansig functions gave generously a sig-
nificantly better performance than softmax. Consequently,
we could use tansig or purelin as activation functions for
the output layer.
•Learning and Training Functions: Recommended
learning and training functions are learngd and trainr.
learngdm andtraingdx also gave similar performance to
learngd and trainr.
It has been seen from the above research between the
testable hypothesis and successful trials that the supposed
hypothesis is right.
B. Second Hypothesis
Performance of Neural Network rises as we increase the
significant number of hidden/deep layers but after a perimeter
performance reduces and if we frequently increase hidden
layers performance surges again. This rise and fall in the
performance of the neural network would persist if we stay
increasing hidden layers.
Experimental results can be positively confirmed in Table II.
Key Parameters which remained constant are as follows:
•MSE = 0.01
•Epochs = 100
•Training Function = trainr
•Learning Function = learngd
•Output Layer Functions = tansig t, purelin p
It has been experienced from the above experiments that
accuracy remains more than 95% in almost all experiments.
An increasing number of deep layers did not affect the desired
accuracy. Our Neural Network gave almost the same accuracy
with the single deep layer and with five active neurons with ten
hidden layers with twenty neurons. This confirms the hypoth-
esis made does not match with real results because there is no
fluctuation in the accuracy rate against various experiments.
TABLE I
EFFIC IEN CY O F SELE CTE D AN D RANDOMLY CHOSE N DATA
Hidden Activation Output Learning Training Accuracy Accuracy
Layers Function Layer
Function
Function Function Selected
Data
(% )
Random
Data
(% )
{20}tansig tansig learngd trainr 97.7213 73.4621
{3×20}tansig tansig learngd trainr 99.0152 74.1921
{5×20}tansig tansig learngd trainr 97.4251 74.6134
{15}logsig purelin learngdm traingdx 98.6157 72.9154
{3×15}logsig tansig learngdm traingdx 97.7622 75.0145
{5×15}logsig softmax learngdm traingdx 34.5462 0.0000
{10}tansig softmax learngd trainr 34.4877 0.0000
{3×10}tansig tansig learngd trainr 98.6231 09.4167
{5×10}tansig purelin learngd trainr 98.9832 73.3917
TABLE II
PERFORMANCE OF NEURAL NET WOR K
Hidden Activation Accuracy Time
Layers Function (% ) (secs)
{20}tansig t = 97.2171
p = 98.3634
t = 08
p = 18
{3×20}tansig t = 96.4486
p = 96.8268
t = 12
p = 09
{5×20}tansig t = 97.9255
p = 97.9912
t = 10
p = 11
{15}logsig t = 96.9229
p = 95.6219
t = 11
p = 16
{3×15}logsig t = 98.8186
p = 97.9857
t = 20
p = 31
{5×15}logsig t = 98.2507
p = 96.9816
t = 34
p = 29
{10}tansig t = 94.6014
p = 98.1173
t=9
p = 11
{3×10}tansig t = 96.9129
p = 97.1429
t = 09
p = 25
{5×10}tansig t = 96.6144
p = 98.9571
t = 35
p = 30
Therefore, the above hypothesis could be correctly stated as
follows.
The performance of the Neural Network is unaffected with
the continual increase of hidden layers. As a result, the
increasing number of hidden layers, after a maximum, would
be of no use.
C. Third Hypothesis
Academic performance of Neural Network progressively
increases if we train Neural Network on fewer-dimensional
data instead of higher-dimensional data.
We know precisely that every unique attribute (symptoms
of a Cell) is mapped to a real number in the domain of 1−10.
If an attribute is closer to 1, it properly indicates benign and if
it is closer to 10 it indicates malignant [2]. If we take a look
at the given data, we comprehend that in most of the specific
cases if the values of the first three columns are greater than
or equal to 5 it is malignant and if the values of the last three
columns are less than 5, it is benign. For training purposes, we
gently separate the first three columns for malignant and last
three columns for benign. In three dimension data, complexity
can be recognized in Fig 3. Red dots accurately indicate
Malignant and blue dots indicate Benign. The key parameters
for this experiment are the same as for the previous hypothesis.
Experimental results are positively discussed in Table III.
TABLE III
PER FOR MA NCE O F FEWE R AN D HIGHER DIMENSIONAL DATA
Hidden Activation Accuracy Time
Layers Function (% ) (secs)
{1×5}1×tansig t = 97.9714
p = 97.9714
t = 29
p = 30
{2×5}2×tansig t = 98.3714
p = 97.9714
t = 37
p = 26
{3×5}3×tansig t = 99.1329
p = 98.4519
t = 60
p = 34
{1×10}1×tansig t = 97
p = 97
t = 28
p = 55
{2×10}2×tansig t = 98.6031
p = 97.9857
t = 24
p = 21
{3×10}3×tansig t = 99.2507
p = 98.9816
t = 32
p = 13
{1×20,2×10}3×tansig t = 99.6014
p = 99.1173
t = 16
p = 33
{2×20,1×10}3×tansig t = 98.9129
p = 99.1429
t = 15
p = 09
{3×20}3×tansig t = 98.6144
p = 98.6144
t = 10
p = 25
Fig. 3. Data Complexity of Benign and Malignant.
It has been observed from above experiments that if we
bring data down to fewer dimensions we achieve more precise
accuracy as compared to higher dimensional data because
all the experiments carried out earlier gave more than 98%
of accuracy which is precisely a more good precision as
compared to experiments performed in the first hypothesis, on
9 dimensional data and some of the experiments gave almost
100% accuracy.
It has been noticed from the above investigation between
the hypothesis and experiments that the supposed hypothesis
is correct.
D. Fourth Hypothesis
With a higher number of hidden layers, Neural Network
takes less time to be trained as compared to the time with
less number of hidden layers. The same criteria are properly
used for the previous hypothesis and all of them are constant
throughout the unique experiment.
Experimental results are reported in Table IV.
TABLE IV
TIM E TAKE N BY NEURAL NET WOR K ON DI FFER EN T EXPE RI MEN TS
Hidden Activation Time
Layers Function (secs)
{1×5}1×tansig t = 22
p = 18
{3×5}2×tansig t = 25
p = 19
{5×5}3×tansig t = 11
p = 10
{1×10}1×tansig t = 11
p = 22
{2×10}2×tansig t = 20
p = 10
{3×10}3×tansig t = 19
p = 17
{1×20,2×10}3×tansig t = 14
p = 21
{2×20,1×10}3×tansig t = 15
p = 18
{3×20}3×tansig t = 19
p = 22
The above-stated hypothesis is incorrect because sometimes
Neural Network takes more time in training on a higher
number of hidden layers and sometimes it takes less time in
training with a lesser number of hidden layers. So, time is
not affected by the number of hidden layers and its size. The
given hypothesis could be stated as:
Time taken by Neural Network during training is
independent of the number and size of hidden layers.
The specific major working of ANN is accurately reported
in Fig 4, Fig 5, and Fig 6 in practical terms of finding and
successful performance.
VI. CONCLUSION
In this academic paper, breast cancer classification and
sufficient proof of key Artificial Neural Network Terminolo-
gies are accurately reported. The key focus is on describing
different ways by which better results can be achieved for
classifying breast cancer. Different hypotheses are brought
into experiments to assess their logical validity and also
discussed the environmental set up to make up an efficient
Artificial Neural Network to positively identify breast cancer.
The overall achievement of the research is precisely 99%.
Furthermore, in an alternative way, sufficient proofs of
some key terminologies traditionally used in ANNs are also
discussed which will clarify the fundamental concepts of
Fig. 4. Validation Performance of ANN
Fig. 5. Validation Checks
ANNs. This will help in establishing key concepts of those
who are new in this academic field as well.
In our future work, the focus will be on Mammography
image processing using AI and other latest approaches for the
detection of malignant and benign.
REFERENCES
[1] “Cancer.org. (2017). How Common Is Breast Cancer?. [online] Available
at: https://www.cancer.org/cancer/breast-cancer/about/how-common-is-
breast-cancer.html,” 2018.
[2] A. Hanikoglu, E. Kucuksayan, F. Hanikoglu, T. Ozben, G. Menounou,
A. Sansone, C. Chatgilialoglu, G. Di Bella, and C. Ferreri, “Effects
of somatostatin, curcumin and quercetin on the fatty acid profile of
breast cancer cell membranes,” Canadian journal of physiology and
pharmacology, no. ja, 2019.
[3] M. Malvezzi, G. Carioli, P. Bertuccio, P. Boffetta, F. Levi, C. La Vecchia,
and E. Negri, “European cancer mortality predictions for the year 2019
Fig. 6. Linear regression
with focus on breast cancer,” Annals of Oncology, vol. 30, no. 5, pp.
781–787, 2019.
[4] 2019. [Online]. Available: https://www.nationalbreastcancer.org/breast-
cancer-facts
[5] E. Alickovic and A. Subasi, “Normalized neural networks for breast
cancer classification,” in International Conference on Medical and
Biological Engineering. Springer, 2019, pp. 519–524.
[6] F. Schnabel, S. Pivo, E. Dubrovsky, J. Chun, S. Schwartz, A. Guth,
and D. Axelrod, “Hormone replacement therapy and breast density after
surgical menopause,” 2018.
[7] M. S. Salama, A. S. Eltrass, and H. M. Elkamchouchi, “An improved
approach for computer-aided diagnosis of breast cancer in digital
mammography,” in 2018 IEEE international symposium on medical
measurements and applications (MeMeA). IEEE, 2018, pp. 1–5.
[8] J. Wang, C. Zhao, C. Shi, S. Tamura, and N. Tomiyama, “A two-
stage high-dimensional feature selection method for pulmonary tumor
classification in ct,” Journal of Medical Imaging and Health Informatics,
vol. 9, no. 7, pp. 1516–1523, 2019.
[9] M. A. Mohammed, B. Al-Khateeb, A. N. Rashid, D. A. Ibrahim,
M. K. A. Ghani, and S. A. Mostafa, “Neural network and multi-fractal
dimension features for breast cancer classification from ultrasound
images,” Computers & Electrical Engineering, vol. 70, pp. 871–882,
2018.
[10] J. Yang and J. Ma, “Feed-forward neural network training using sparse
representation,” Expert Systems with Applications, vol. 116, pp. 255–
264, 2019.
[11] J. Singh and R. Banerjee, “A study on single and multi-layer perceptron
neural network,” in 2019 3rd International Conference on Computing
Methodologies and Communication (ICCMC). IEEE, 2019, pp. 35–40.
[12] S. Dutta, X. Chen, S. Jha, S. Sankaranarayanan, and A. Tiwari,
“Sherlock-a tool for verification of neural network feedback systems:
demo abstract,” in Proceedings of the 22nd ACM International Confer-
ence on Hybrid Systems: Computation and Control. ACM, 2019, pp.
262–263.
[13] J. I. Glaser, A. S. Benjamin, R. Farhoodi, and K. P. Kording, “The roles
of supervised machine learning in systems neuroscience,” Progress in
neurobiology, 2019.
[14] N. Andrei, “A dai-liao conjugate gradient algorithm with clustering of
eigenvalues,” Numerical Algorithms, vol. 77, no. 4, pp. 1273–1282,
2018.
[15] N. M. Nawi, N. H. M. Saufi, A. Budiyono, N. A. Hamid, M. Z. Rehman,
and A. A. Ramli, “An improved back propagation leaning algorithm
using second order methods with gain parameter,” International Journal
of Integrated Engineering, vol. 10, no. 6, 2018.
[16] A. Mujika, F. Meier, and A. Steger, “Approximating real-time recur-
rent learning with random kronecker factors,” in Advances in Neural
Information Processing Systems, 2018, pp. 6594–6603.
[17] C. He, Y. Liu, T. Yao, F. Xu, Y. Hu, and J. Zheng, “A fast learning
algorithm based on extreme learning machine for regular fuzzy neural
network,” Journal of Intelligent & Fuzzy Systems, no. Preprint, pp. 1–7,
2019.
[18] S. Kim, S. Ahn, S. Leem, J. Jeong, and I. Chu, “116p genomic
characteristics of standardized uptake value of 18f-fluorodeoxy-glucose
positron emission tomography in breast cancer,” Annals of Oncology,
vol. 29, no. suppl 8, pp. mdy269–114, 2018.
[19] K. S. Alexander, C. Baker, P. Smith, and D. Grinsell, “The combination
single ct scan for breast cancer staging and reconstruction,” Annals of
Breast Surgery, vol. 2, no. 4, 2018.
[20] J. Deng, S. Xu, W. Hu, X. Xun, L. Zheng, and M. Su, “Tumor
targeted, stealthy and degradable bismuth nanoparticles for enhanced
x-ray radiation therapy of breast cancer,” Biomaterials, vol. 154, pp.
24–33, 2018.
[21] R. Sood, A. F. Rositch, D. Shakoor, E. Ambinder, K.-L. Pool, E. Pollack,
D. Mollura, L. A. Mullen, and S. C. Harvey, “Handheld ultrasound for
breast cancer detection in low-resource settings: A systematic review
and meta-analysis,” Available at SSRN 3362448, 2019.
[22] E. L. Henriksen, J. F. Carlsen, I. M. Vejborg, M. B. Nielsen, and C. A.
Lauridsen, “The efficacy of using computer-aided detection (cad) for
detection of breast cancer in mammography screening: a systematic
review,” Acta Radiologica, vol. 60, no. 1, pp. 13–18, 2019.
[23] A. Al-Khasawneh, “Diagnosis of breast cancer using intelligent infor-
mation systems techniques,” in Nature-Inspired Computing: Concepts,
Methodologies, Tools, and Applications. IGI Global, 2017, pp. 203–
214.
[24] G. Villarrubia, J. F. De Paz, P. Chamoso, and F. De la Prieta, “Artificial
neural networks used in optimization problems,” Neurocomputing, vol.
272, pp. 10–16, 2018.
[25] S. Huang, N. Cai, P. P. Pacheco, S. Narrandes, Y. Wang, and W. Xu,
“Applications of support vector machine (svm) learning in cancer
genomics,” Cancer Genomics-Proteomics, vol. 15, no. 1, pp. 41–51,
2018.
[26] V. V. Spencer, A Framework for Improving Breast Cancer Care Deci-
sions by Using Self-Organizing Maps to Profile Patients and Quantify
Their Attributes. Mississippi State University, 2018.
[27] M. Kusy and P. A. Kowalski, “Weighted probabilistic neural network,”
Information Sciences, vol. 430, pp. 65–76, 2018.
[28] D. Bani-Hani, P. Patel, and T. Alshaikh, “An optimized recursive general
regression neural network oracle for the prediction and diagnosis of
diabetes,” Global Journal of Computer Science and Technology, 2019.
[29] P. Zarbakhsh, A. Addeh et al., “Breast cancer tumor type recognition
using graph feature selection technique and radial basis function neural
network with optimal structure,” Journal of cancer research and thera-
peutics, vol. 14, no. 3, p. 625, 2018.
[30] W. H. Land and J. D. Schaffer, “The generalized regression neural
network oracle,” in The Art and Science of Machine Intelligence.
Springer, 2020, pp. 77–105.
[31] H. Chen, B. Jiang, and X. Yao, “Semisupervised negative correlation
learning,” IEEE transactions on neural networks and learning systems,
vol. 29, no. 11, pp. 5366–5379, 2018.
[32] Z. Shi, L. Zhang, Y. Liu, X. Cao, Y. Ye, M.-M. Cheng, and G. Zheng,
“Crowd counting with deep negative correlation learning,” in Proceed-
ings of the IEEE conference on computer vision and pattern recognition,
2018, pp. 5382–5390.
[33] 2019. [Online]. Available: https://archive.ics.uci.edu/ml/datasets/Breast+
Cancer+Wisconsin+(Diagnostic)
