Breast Cancer Diagnosis and Comparative Studies

using effective Deep Neural Network

1st Nazmin Islam 2nd Oishi Jyoti 3rd Fahim MD. Sifnatul Hasnain
Dept. of Electrical & Dept. of Electrical & Dept. of Electrical &
Computer Engineering, Computer Engineering, Computer Engineering,
Rajshahi University of Rajshahi University of Rajshahi University of
Engineering & Technology, Engineering & Technology, Engineering & Technology,
Rajshahi, Bangladesh Rajshahi, Bangladesh Rajshahi, Bangladesh
[email protected] [email protected] [email protected]

Abstract—Advance prognosis of any disease can be treatable in middle income countries and below 40% in low-income
with a small proportion of effort. Most of the people around the countries [2]. Moreover, Human life really relies on medical
world are unable to detect their disease before it becomes im- data to detect and predict the disease properly. Datasets are
medicable. It conducts an extension in the death rate widespread.
Breast cancer is the second supreme root of deaths in women used for obtaining the remedy for those diseases so that the
worldwide and happens in almost one out of eight women. It is disease can be cured properly.
one of the diseases that could be healed only when the disease Breast cancer has many signs like Cell Size, Cell Shape,
is recognized at a premature phase prior to it unrolling over all Clump Thickness, Bare Nuclei, compactness, smoothness and
of the organs of the body. The medical specialists may diagnose many more. For detecting whether the cancer is benign or
the diseases erroneously on account of misapprehend. All these
reasons are a complete sign of detecting the disease properly as malignant properly, a standard dataset of breast cancer has
expeditiously as possible. By scrutinizing the work of Deep Neural been collected to differentiate between both of two. A better
Network (DNN) models, this will help the doctors to anticipate model will be produced to predict breast cancer which will
breast cancer and its symptoms successfully and only then it further help to detect the disease in a quick time and proper
will be possible to furnish the patients with better treatment. action can be taken to cure it.
The chief goal of this work is to build a standard DNN model
that will provide better accuracy in predicting breast cancer. The Diagnosis of disease is one of the major complications in the
dataset applied for this purpose is collected from the University of medical sector. The development of neural networks have been
Wisconsin. To predict whether the cancer is benign or malignant, done by several research groups working for the improvement
the accuracy attained by this model is 98.59% which is finer than of medical diagnosis. In recent years, deep neural networks
the pre-existing work. have been applied to a number of prediction problems. The aim
Index Terms—deep neural network(DNN), breast cancer, clas-
sification, breast cancer prediction. of the system is to develop a novel deep neural network model
to predict if breast cancer is benign or malignant with more
accurate results to save precious lives. DNN helps to generate
I. I NTRODUCTION
a better model to train the given dataset and also supports
Data is exceedingly important in today’s time for mak- algorithms like supervised, unsupervised, reinforcement etc.
ing important decisions which is also the ultimate valuable In this study, a standard DNN model is initiated that provides
resource for any management. Creation of data is the most an accuracy of 98.59% to predict the cancer and in addition
important thing in the world right now which occupies every a better knowledge of working methods of various functions
sector of our life like business, medicine, industrial etc. Breast can be gained.
cancer is one of the leading cancers for women around The remaining portion of the paper provides the chapters
the world. The complexities of breast cancer are inflating which is summarized as follows: Section 2 describes the
in the world due to adoption of extremely busy lifestyles, related or previous work. In Section 3 the proposed method
relocation of people from rural to urban inhabitants, increased is explained in more detail. Section 4 demonstrates the ex-
life anticipation and many other reasons. Because of these perimental results and discussion. Section 5 concludes the
reasons, people certainly are not aware of these diseases and conclusion.
they do not care too. Breast cancer is the most perennial cancer
among females, affecting 2.1 million women every year, and II. R ELATED W ORK
further originates the biggest number of cancer-related deaths. Recently, there is a lot of work going on predicting the
In 2018, it is estimated that 627,000 women died from disease correctly in earlier stages. Savita Kumari [3] inves-
breast cancer – that is approximately 15% of all cancer deaths tigated the Wisconsin Breast Cancer Data (WBCD) through
among women [1] . Statistics show that ranging from 80% the Big Data approach using the Hybrid SVM RVM Model as
or over in North America, Sweden, and Japan to around 60% a classifier and got 96.41% accuracy. In paper [4], Adaptive
Resonance Theory Based NN (ART), Self Organizing Map- to establish a relationship present in the brain followed by
Based NN (SOM) and Back Propagation NN (BPN) are used a homogeneous artificial neural network with a complex
for the Wisconsin breast cancer database and compare with network composition which has ‘n’ hidden layers. Deep Neural
processing speed and accuracy. Both cases BPN gives the Network (DNN) is the recent remarkable computational model
best result with 91% accuracy. Compared with the other two inspired by the biological human brain to solve complex
models, ART gives poor accuracy. Mert et al. [5] used a problems using the stacked layers of artificial neurons in a
radial basis function neural network (RBFNN) for medical deeper manner [13]. Moreover, DNN can operate the input
data classification and independent component analysis for data from the previous layer. By just modifying weights of
feature selection. The method selects the one feature vector each node, the error rate of the input data can be reduced
randomly from 30 features. The method obtained an accuracy constantly and this leads to attain the goal of achieving more
of 86%. In paper [6], a more cost-effective and easy–to-use accurate results. The system need not to define any specific
system for supporting clinicians is developed. For the breast algorithm as it defines complex hierarchies in a simple form.
cancer tumor diagnosis problem, experimental results show The DNN creates a better model themselves to train the
that the concise models extracted from the network achieve given data. It continues to backpropagate the network until
a high accuracy rate on the training data set and on the test and unless it gains a better result. Deep neural networks
data set. Studies done in [7] illustrate that different parts of use sophisticated mathematical modeling to process data in
SVM, ANN and Naive Bayes are trained based on a particular complex ways. It uses multiple deep layers of units with highly
training dataset (WBCD). Based on the highest accuracy, the optimized algorithms and architectures [14].
best model has been voted from the several models like linear
SVM, support vector machine, Feedforward and Radius basis
function in this paper and selected to use in the future for the
client dataset (clinical data).
Schmidhuber [8] provides an overview of deep learning in
neural networks. The method proves that the deep learning
algorithms reduced the error rate and increased the accuracy
with respect to training of algorithms. Sridevi et al. [9]
classified the Breast Cancer dataset using K-Means clustering
to select the subsets of features and then applied the SVM
algorithm to classify the dataset. Jouni et al. [10] proposed a
model based on artificial neural networks with multilayered
perceptron networks and BPNN. This model learns to classify
whether the result of the simulation will be malignant or
benign. It also includes weight adjustment factors and bias
values. Bewal et al. [11] used multilayered perceptron net-
work with four back-propagation training algorithms such as Fig. 1. Architecture of Deep Neural Network
quasi-Newton, gradient descent with momentum and adaptive
learning, Levenberg–Marquardt, and resilient backpropagation Any number of inputs can be used in the input layer. Again,
to train the network. Steepest descent back propagation is to increase the process of learning, the numbers of nodes are
used to measure the performance of other neural networks. more than the input layer diligently. In the output layer,the
Levenberg– Marquardt algorithm with MLP achieved best number of outputs can be defined separately and the output
accuracy rate of 94.11 nodes are defined uniquely. The variables that are used in DNN
are said to be bias, learning rate, initial weights, number of
III. D EEP N EURAL N ETWORK
nodes in input, output and hidden layer and stop condition for
In today’s world, neural networks are one of the rising fields concluding the epochs. Normally, bias value is assigned as 1
that occupy every scope in the world and also they are the most in any neural network only to avoid null results in the network.
promising one in terms of work. The painstakingly handcrafted Initially, the weights of the nodes can be allocated randomly.
feature extractors used in traditional learning, classification, Based on the number of inputs and sizes, the number of
and pattern recognition systems are not scalable for large- nodes and hidden layers are decided.Termination condition of
sized data sets. In many cases, depending on the problem the network depends on the epochs achieved by the learning
complexity, DL can also overcome the limitations of earlier model. Depending on the number of nodes and layers, the
shallow networks that prevented efficient training and ab- execution time varies.A constructive architecture of DNN is
stractions of hierarchical representations of multi-dimensional shown above in Fig. 1.
training data [12]. Neural Networks are used to transfer data
in the form of input and output values by making connections IV. DATABASE D ESCRIPTION
thoroughly where deep neural networks are associated with the In this study, a standard dataset is used to predict the breast
transformation and extraction of features which endeavours cancer is benign or malignant containing features which are
computed from a digitized image of a Fine Needle Aspirate A. Dataset Analysis
(FNA) of a breast mass that explain the properties of the cell Processed and balanced dataset assist to get better accuracy
nuclei present in the image. The data occupies a total of 570 in- over the output.A complete data set consisting of classical
stances of patients from the University Of Wisconsin Hospitals influence factors measured at time of primary therapy together
and no patients with missing data has been found. The features with follow-up data (time of relapse or censoring) was used for
which are real-valued are calculated for every nucleus cell. The the analysis [17]. In order to evaluate the proposed model, the
database contains 32 columns representing attributes which breast cancer dataset has been processed to meet the expected
are radius, texture, perimeter, area, smoothness, compactness, accuracy and precision. Around 570 features are computed
concavity, concave points, symmetry and fractal dimension. from a digitized image of a fine needle aspirate of a breast
The mean, standard error and worst or largest that means mean mass. In total 32 features are bounded in the dataset. Patients’
of the three largest values of these features were computed for identity information (ID) has been dropped as it is a unique
every image, resulting in 30 features. For example, column value and does not affect the output of the disease. The
3 is Mean Radius, column 13 is Radius SE, column 23 is features give general information and prefixes related to the
Worst Radius. All the values of the attributes are documented breast cancer including radius, texture, perimeter, smoothness,
with four significant digits that represent the data type float. compactness, concavity etc. The feature columns are subject
The last column occupies a class attribute named ‘Diagnosis’ to mean value, standard error value and largest mean value of
that shows the output result representing either benign(non- the referred parameters. The data were converted in numeric
cancerous) or malignant(cancerous). value including floating point and integer point format as the
V. M ETHODOLOGY deep learning operates with numeral data only. Consequently
the last feature attribute ‘Diagnosis’ was ordained as 0 for
In recent years, deep neural networks (including recurrent having Malignant (M) and 1 for Benign (B). The value of
ones) have won numerous contests in pattern recognition and the features are well balanced and impartial which guides the
machine learning [15]. Deep Neural Network is a promising algorithm to look for maximizing accuracy of all classes in a
field in biomedical engineering, computational intelligence, more balanced way.
data science etc. as it is capable of finding the efficient
mathematical way to convert the input into the output which
builds the enormous accuracy and precision. Neural networks
are trained to perform specific functions by adjusting weights
between elements [16]. In this proposed paper, DNN has been
used to train and evaluate the model.

Fig. 3. Representation of some of the features from the dataset

B. Case Training
Training and development models bestow insight, formation
and productivity for specific initiatives that align with the
model. Training design models contain five steps which are
analysis, design, development, implementation and evaluation.
To be able to evaluate the performance of various classification
methods, there is a need to introduce quantitative criteria [18].
In this paper, a deep neural network model has been applied to
inspect the patients’ tumor whether it is malignant or benign
and intuites about breast cancer. The network was trained
using back propagation algorithm [19]. It was instructed and
evaluated using standard 50%-50% deviation that is 50% was
used as training data which is 285 from data and 50% was
used as data testing data which is 285 from the dataset.
Training process smeared two different hidden layers. For a
sophisticated result, 12 nodes were delegated to the first hidden
Fig. 2. A schematic representation of the Methodology layer and the next one holds 12 nodes. Initializing the weights
for the hidden layer glorot uniform was used. Mean Squared
Error (mse) was assigned to calculate the loss. The model was 97.54%. During the training process, the adopted activation
trained for 10,000 epochs with 0.1 learning rate. To inspect functions enable the network to estimate a better output by
the performance batch training approach was applied to sustain priming the activation function parameters.
the model through training data. When the trained network is
tested using some testing data values then it generates the D. Assessment Parameters
output.The trained network gives the output which is either In order to demonstrate the operation of the classification
malignant or benign [20]. algorithms, the confusion matrix was used from where the
accuracy, precision, recall, f1-score for every model can be
C. Performance measure of Activation Function calculated. A confusion matrix contains information about
In a deep neural network, the activation function is ac- actual and predicted classifications done by a classification
countable for converting the added weighted input from the system [21]. Elements of an actual class are constituted by
node into the output for that input or activation of the node. each row of the confusion matrix and the instances of a
The objective is to include some kind of irregular property to predicted class is constituted by each column. The best model
the function. Language transformation, image classification, for the Breast Cancer dataset can be achieved by utilizing the
detection of objects etc are very intricate in deep learning result of the accuracy for every model.
which are needed for inscriptions with the help of activation
function. It helps in determining the output of deep learning Table-2 : Confusion Matrix
models, accuracy and performance efficiency of the training
model that can design or divide a huge scale neural network.
Enhancing the standardization of the network to deal with
the several applications is operated by activation function.
Activation functions have considerable consequences on the
potentiality of neural networks to converge and convergence
speed. In this paper, DNN performance is explored with After iteration by 10,000 epoch and 0.1 learning rate the
hyperbolic tangent, sigmoid, softplus, rectified linear units and achieved confusion matrix evaluating the effectual classifica-
swish activation function. tion operation for the model for 98.59% accuracy is
 
105, 4
Table-1 : Accuracy Comparison with Activation
0, 176
Functions
All the models were simulated using a variety of parameters
[22].To measure performance accuracy and precision for every
algorithm activation function is used. Accuracy is the mea-
surement of value to a standard or true value. precision is
delegated for exploring the adjustment or reproducibility of
the assessment. Recall is the fragment of the total amount
of pertinent features that were actually redeemed. Moreover,
these parameters assist the training performance competently.
Fig.4 illustrates the comparison of different accuracy at- The evaluation parameters are illustrated below [23].
tained from activation functions.

In these formulas, TP, TN, FP and FN represent the true

Fig. 4. Train Data Comparative Accuracy Curve For Different Activation
Functions positive number, the true negative number, the false positive
number and the false negative number, respectively [24]. For
The experimental results showed that sigmoid activation func- the 98.59% accuracy, inspecting the confusion matrix attained
tion unveiled the maximum accuracy with 10K epochs and evaluation parameters are given below.
0.1 learning rate which is 98.59%. Furthermore, hyperbolic The computed accuracy is alike with the testing accuracy
tangent activation function obtained the lowest accuracy with whis is apparent. Evaluation parameter facilitates the way to
 Fig. 5. Comparative Loss Curve
obtain the ROC (Receiver Operating Characteristic) curve to
evaluate the implementation of the model that evinces graph-
ical representation. The ROC curve summarizes the trade-off represents the comparative loss curve of the experimented
between the true positive rate and false positive rate for a model for the same epochs and learning rate as before.
model using different probability thresholds [25]. Here this is the Receiver Operating Characteristic curve
(ROC) in Fig.6 exemplifies the diagnostic potentiality of
E. Environment binary classifiers and also gives an idea about the tested model.
Attaining the contemplated goal of better accuracy in pre-
diction for breast cancer is the most significant substance
of this paper. To accomplish this fact, the environment for
performing the network plays an important role. A suitable
environment will adjust the prediction model as expected.
Python-based ecosystem of open-source software has been
used in this paper. In addition, numpy, pandas, sci-kit learn,
matplotlib, etc. have been combined with keras to enable as a
backend to build our model in processing the data. Matplotib
was used to create animated, static and collaborative visual-
izations. The Scikit-Learn library reinforced by the Python
programming language was delegated to split the dataset, pre- Fig. 6. Receiver Operating Characteristic curve (ROC)
process and select models effectively. To numerically compute
and build models using data flow graphs Tensorflow was used The experimented results signify the better performance
which is an open source artificial intelligence library. Large- of deep neural networks that helps people worldwide to
scale neural networks with many layers were created through diagnose the severe breast cancer disease more accurately with
this. DNN model, flow diagram, confusion matrix was built a minimum amount of time that makes the model efficiently
using PowerPoint. To evaluate parameters Mathtype software valuable and effective.
was used. Google Colaboratory was chosen for performing
programming because of its efficiency. VII. C ONCLUSION
In this modern time, people have to face modern age
VI. E XPERIMENTAL R ESULTS AND D ISCUSSION diseases. One of the most common types of deadliest disease
DNN plays a vital role in medical diagnosis with becoming raising over time among different countries is breast cancer.
the promising one to give better accuracy with better perfor- Paucity of apprehension and post-recognition of disease is the
mance. Various activation functions have been experimented crucial reason for more death rates than before. There is a
throughout this research to acquire the best result for the proverb that prevention is better than cure which indicates the
patients who suffer from this chronic disease and also for necessity of keeping bad things from happening and trying an
the medical diagnosis to get the outcome more perfectly and effort to fix that thing because every life matters. At the same
quickly. From the experimented result, it is comprehensible time, due to the rapid increase of these diseases, it also helps
that deep neural networks give better accuracy with better the doctors to predict the disease more efficiently and easily.
performance in predicting breast cancer remarkably. Computer-aided diagnosis will be a great solution for people
As shown in Table-1, among all activation functions,five to diagnose with accurate results among which deep neural
of them are experimented with a 0.1 learning rate and 10k networks are the recent top model to work with as it gives more
epochs for all. Among all of them, the highest accuracy that accuracy and its performance makes it preferable for all. Due
has been achieved by the sigmoid function through this process to so many problems in the medical system, the best accuracy
is 98.59%. needs to be provided to make the best decision for the patients
Accuracy and loss curve are the indication of performance as early as possible. In this study, a standard DNN model
evaluation of delegated activation functions. Therefore, Fig.5 has been presented for detecting breast cancer is benign or
malignant. An accuracy of 98.59% was achieved successfully [19] Saritas, I. Prediction of Breast Cancer Using Artificial Neural Networks.
using a deep neural network. Evaluation of activation functions J Med Syst 36, 2901–2907 (2012). https://doi.org/10.1007/s10916-011-
9768-0
helps to give more accuracy with minimum loss that makes [20] P. Singhal and S. Pareek, ”Artificial Neural Network for Prediction of
the proposed system more reliable than the previous models Breast Cancer,” 2018 2nd International Conference on I-SMAC (IoT
and also the final outcome of the model benefits the user of in Social, Mobile, Analytics and Cloud) (I-SMAC)I-SMAC (IoT in
Social, Mobile, Analytics and Cloud) (I-SMAC), 2018 2nd International
this system who will definitely obtain a computationally better Conference on, Palladam, India, 2018, pp. 464-468, doi: 10.1109/I-
solution to test and detect the result. SMAC.2018.8653700.
[21] R. R. Janghel, A. Shukla, R. Tiwari and R. Kala, ”Breast cancer
diagnosis using Artificial Neural Network models,” The 3rd International
R EFERENCES Conference on Information Sciences and Interaction Sciences, Chengdu,
2010, pp. 89-94, doi: 10.1109/ICICIS.2010.5534716.
[1] P. Basu, F. Selmouni, L. Belakhel, C. Sauvaget, L. Abousselham, [22] Janghel, Rekh Shukla, A. Tiwari, Ritu Kala, Rahul. (2010). Breast
E. Lucas, R. Muwonge, R. Sankaranarayanan, and Y. C. Khazraji, Cancer diagnosis using Artificial Neural Network models. Proceedings
“Breast cancer screening program in morocco: Status of implementation, - 3rd International Conference on Information Sciences and Interaction
organization and performance,” International journal of cancer, 2018. Sciences, ICIS 2010. 89 - 94. 10.1109/ICICIS.2010.5534716.
[2] M. P. Coleman, M. Quaresma, F. Berrino, J.-M. Lutz, R. De Angelis, [23] Sekaran, Karthik Ramalingam, Srinivasa Perumal P.V.S.S.R., Chandra
R. Capocaccia, P. Baili, B. Rachet, G. Gatta, T. Hakulinen et al., Mouli. (2018). Breast Cancer Classification Using Deep Neural Net-
“Cancer survival in five continents: a worldwide population-based study works. 10.1007/978-981-10-6680-1-12.
(concord),” The lancet oncology, vol. 9, no. 8, pp. 730–756, 2008. [24] A. K. Arslan, Ş. Yaşar and C. Çolak, ”Breast cancer classification
[3] S. K. Sheoran, “Breast cancer classification using big data approach,” using a constructed convolutional neural network on the basis of
PARIPEX-INDIAN JOURNAL OF RESEARCH, vol. 7, no. 1, 2018. the histopathological images by an interactive web-based interface,”
[4] K. Mumtaz, S. Sheriff, and K. Duraiswamy, “Evaluation of three 2019 3rd International Symposium on Multidisciplinary Studies and
neural network models using wisconsin breast cancer database,” in 2009 Innovative Technologies (ISMSIT), Ankara, Turkey, 2019, pp. 1-5, doi:
International Conference on Control, Automation, Communication and 10.1109/ISMSIT.2019.8932942.
Energy Conservation. IEEE, 2009, pp. 1–7. [25] Wu, Nan, Jason Phang, Jungkyu Park, Yiqiu Shen, Zhe Huang, Masha
[5] Mert, A., Kılıc¸, N., Bilgili, E., Akan, A. (2015). Breast cancer detection Zorin, Stanislaw Jastrzebski, Thibault Févry, Joe Katsnelson, Eric Kim,
with reduced feature set. Computational and Mathematical Methods in Stacey Wolfson, Ujas Parikh, Sushma Gaddam, Leng Leng Young Lin,
Medicine. Kara Ho, Joshua D. Weinstein, Beatriu Reig, Yiming Gao, Hildegard
[6] Muhammad Sufyian Bin Mohd Azmi and Zaihisma Che Cob, ”Breast Toth, Kristine Pysarenko, Alana Lewin, Jiyon Lee, Krystal Airola, Eralda
Cancer prediction based on Backpropagation Algorithm,” 2010 IEEE Mema, Stephanie Chung, Esther Hwang, Naziya Samreen, S. Gene Kim,
Student Conference on Research and Development (SCOReD), Putra- Laura Heacock, Linda Moy, Kyunghyun Cho and Krzysztof J Geras.
jaya, 2010, pp. 164-168, doi: 10.1109/SCORED.2010.5703994. “Deep Neural Networks Improve Radiologists’ Performance in Breast
[7] M. I. H. Showrov, M. T. Islam, M. D. Hossain and M. S. Ahmed, Cancer Screening.” IEEE Transactions on Medical Imaging 39 (2020):
”Performance Comparison of Three Classifiers for the Classification of 1184-1194.
Breast Cancer Dataset,” 2019 4th International Conference on Elec-
trical Information and Communication Technology (EICT), Khulna,
Bangladesh, 2019, pp. 1-5, doi: 10.1109/EICT48899.2019.9068816.
[8] Schmidhuber, J. (2015). Deep learning in neural networks: An overview.
Neural Networks, 61, 85–117.
[9] T. Sridevi and A. Murugan, “An intelligent classifier for breast cancer
diagnosis based on k-means clustering and rough set,” International
Journal of Computer Applications, vol. 85, no. 11, 2014.
[10] Jouni, H., Issa, M., Harb, A., Jacquemod, G., Leduc, Y. (2016). Neural
network architecture for breast cancer detection and classification. In:
IEEE International Multidisciplinary Conference on Engineering Tech-
nology (IMCET), pp. 37–41.
[11] Bewal, R., Ghosh, A., Chaudhary, A. (2015). Journal of Clinical and
Biomedical Sciences, 5(4), 143–148.
[12] A. Shrestha and A. Mahmood, ”Review of Deep Learning Algorithms
and Architectures,” in IEEE Access, vol. 7, pp. 53040-53065, 2019, doi:
10.1109/ACCESS.2019.2912200.
[13] Lau, Mian Lim, King Hann. (2018). Review of Adaptive
Activation Function in Deep Neural Network. 686-690.
10.1109/IECBES.2018.8626714.
[14] Ruder, Sebastian. ”An overview of gradient descent optimization algo-
rithms.” arXiv preprint arXiv:1609.04747 (2016).
[15] Schmidhuber, J. (2015).” Deep learning in neural net-
works: An overview. Neural Networks”, 61, 85–117.
doi:10.1016/j.neunet.2014.09.003
[16] Tariq, Nadeem. (2018). “Breast Cancer Detection using Artificial
Neural Networks. Journal of Molecular Biomarkers Diagnosis.” 09.
10.4172/2155-9929.1000371.
[17] R.Kates, N. Harbeck and M. Schmitt, ”Prospects for clinical decision
support in breast cancer based on neural network analysis of clinical sur-
vival data,” KES’2000. Fourth International Conference on Knowledge-
Based Intelligent Engineering Systems and Allied Technologies. Pro-
ceedings (Cat.No.00TH8516), Brighton, UK, 2000, pp. 764-767 vol.2,
doi: 10.1109/KES.2000.884158.
[18] Zemouri, R., Omri, N., Morello, B., Devalland, C., Arnould, L., Zer-
houni, N., Fnaiech, F. (2018). Constructive Deep Neural Network
for Breast Cancer Diagnosis. IFAC-PapersOnLine, 51(27), 98–103.
doi:10.1016/j.ifacol.2018.11.660