Comparison of Machine Learning Methods for Breast

Cancer Diagnosis

Ebru Aydındag Bayrak, Pınar Kırcı Tolga Ensari

Department of Engineering Sciences Department of Computer Engineering
Istanbul University Cerrahpasa Istanbul University Cerrahpasa
stanbul, Turkey stanbul, Turkey
[email protected], [email protected] [email protected]

Abstract— Cancer is the common problem for all people in Wisconsin Breast Cancer was used as a dataset and Weka
the world with all types. Particularly, Breast Cancer is the most software was used as a Machine Learning tool. The key
frequent disease as a cancer type for women. Therefore, any performance parameters of machine learning classifiers have
development for diagnosis and prediction of cancer disease is been compared according to accuracy, recall, precision and
capital important for a healthy life. Machine learning techniques ROC area. They have suggested that BN has the best
can make a huge contribute on the process of early diagnosis and performance according to recall and precision values and RF
prediction of cancer. In this paper, two of the most popular
technique has optimum performance in term of ROC area [5].
machine learning techniques have been used for classification of
Wisconsin Breast Cancer (Original) dataset and the classification Ahmad et al. have exercised machine learning algorithms for
performance of these techniques have been compared with each predicting the rate of two years recurrence of breast cancer
other using the values of accuracy, precision, recall and ROC disease. The dataset has been obtained from Iranian Center of
Area. The best performance has been obtained by Support Breast Cancer (ICBC) program, collected the time period of
Vector Machine technique with the highest accuracy. 1997-2008 years. The dataset is consisted of population
characteristics and 22 input variables also the cases have been
Keywords— machine learning; breast cancer; classification; collected from 1189 women of diagnosed breast cancer.
early diagnosis. Artificial Neural Network (ANN), Support Vector Machine
(SVM) and Decision Tree (DT) have been applied and SVM
I. INTRODUCTION has been showed the best performance with highest accuracy
and least error rate [6].
Cancer is the second reason of human death all over the
Bektas and Babur have studied on diagnosis of breast cancer
world and accounts for roughly 9.6 million deaths in 2018.
using machine learning techniques. Kent Ridge Microarray
Globally, for 1 human death in 6 can be said that is caused by
has been used 2 datasets and support vector machine, k-star,
cancer. Almost 70 percent of the deaths from cancer disease
random forest algorithm and voted perceptron have been
happen in countries that have low and middle income [1]. The
applied. Random forest algorithm has been showed more
most common cancer type among women are breast, lung and
performance than applied feature selection method [7].
colorectal, which totally symbolize half of the all cancer cases.
Chen et al. have applied Support Vector Machine
Also, breast cancer is responsible for the thirty percent of all
classification algorithm on Wisconsin Diagnostic Breast
new cancer diagnoses in women [2]. Machine learning (ML)
Cancer dataset. In the study, the training and testing sets have
methods ensure analyzing the data and extracting key
been split as 50-50%, 70-30% and 80-20%. According to
characteristics of relationships and information from dataset.
different training/testing percent, accuracy values have been
Also, it creates a computational model for best description of
calculated [8].
the data. Especially, according to in researches about cancer
In this paper, as SVM and ANN two of the most popular
disease, it can be said that ML techniques can be handled on
machine learning techniques are applied on Wisconsin Breast
early detection and prognosis of cancer [3]. Asri et al. have
Cancer (Original) dataset and the result of applied machine
compared some machine learning algorithms for the risk
learning (ML) techniques are compared according to
prediction and diagnosis of breast cancer. Support Vector
performance metrics. The rest of the paper is arranged as
Machine (SVM), k-Nearest Neighbors (kNN), Naive Bayes
follows: Section II explains used material. Section III
(NB) and Decision Tree (C4.5) have been applied Wisconsin
describes the fundamental points of machine learning
Breast Cancer (Original) dataset. SVM classification method
techniques. The results are explained in Section IV. The paper
has been given the highest accuracy value (97.13 %) with least
ends in conclusion by proposing concluding remarks.
error rate when the experimental results were compared [4].
Bazazeh and Shubair have investigated the comparative study II. DATA SET
of machine learning techniques as Support Vector Machine
(SVM), Random Forest (RF) and Bayesian Network (BN) for In this study, Wisconsin Breast Cancer (Original) (WBC)
detection and diagnosis of breast cancer. The Original dataset is acquired from UCI Machine Learning Repository
and analyzed. The used dataset is consisting of 699 instances

978-1-7281-1013-4/19/$31.00 ©2019 IEEE

that were classified as benign and malignant. Also, dataset has Perceptron and Voted Perceptron are used as ANN classifier
11 integer-valued attributes [9]. in Weka.
The common features of dataset are; The results of the applied machine learning techniques
• Benign class: 458 (65.5 %) application on WBC dataset are reported. We applied k=10-
• Malignant class: 241 (25.5%) fold cross validation and percentage split (% 66 and % 33
The 11 integer-valued attributes of dataset are; splits) which are training options that split the dataset into a
 Sample code number training set to train model and a testing set to evaluate it. In
 Clump Thickness (1-10) WEKA software, SVM and ANN techniques are applied on
 Cell Size Uniformity (1-10) breast cancer dataset. The experimental results are
 Cell Shape Uniformity (1-10) demonstrated in Table 1 and Table 2 respectively. In the
 Marginal Adhesion (1-10) results, performance metrics of accuracy, precision, recall and
 Single Epithelial Cell Size (1-10) ROC Area are compared.
 Bare Nuclei (1-10)
 Bland Chromatin (1-10) Table 1: The experimental results of Support Vector Machine
 Normal Nuclei (1-10) classification techniques.
 Mitoses (1-10) Support The results of performance
Vector ROC
 Class (2: Benign, 4: Malignant) Machine Test Options Accuracy Precision Recall
Area
Sequential k=10 cross
III. EXPERIMENTAL ANALYSIS Minimal validation
0,969957 0,97 0,97 0,968

In this paper, we have applied SVM and ANN techniques for Optimization Percentage
(SMO) 0,953782 0,954 0,954 0,949
split
prediction of the classification of breast cancer to find which k=10 cross
machine learning methods performance is better. 0,957082 0,96 0,957 0,964
validation
LibSVM
Support Vector Machines (SVMs) have been first explained Percentage
0,957983 0,958 0,958 0,957
by Vladimir Vapnik and the good performances of SVMs have split
been noticed in many pattern recognition problems. SVMs can
indicate better classification performance when it is compared Table 2: The experimental results of Artificial Neural Network
with many other classification techniques [10]. SVM is one of classification techniques.
Artificial The result of performance
the most popular machine learning classification technique Neural ROC
that is used for the prognosis and diagnosis of cancer. Network Test Options Accuracy Precision Recall
Area
According to SVM, the classes are separated with hyperplane k=10 cross
Multi-Layer 0,95442 0,954 0,954 0,988
that is consisted of support vectors that are critical samples Perceptron
validation
from all classes. The hyperplane is a separator that is (MLP) Percentage split 0,953782 0,955 0,954 0,994
identified as decision boundary among the two sample k=10 cross
0,909871 0,919 0,91 0,929
clusters. SVM can be used for classifying tumors as benign or Voted validation
Perceptron
malignant based on patient’s age and tumors size [11]. Percentage split 0,882353 0,899 0,882 0,915
Artificial Neural Network (ANN) can be expressed in terms of
biological neuron system. Especially, it is similar to human
brain process system. It is consisted of a lot of nodes that IV. CONCLUSION
connect each node [12]. ANN have the ability of modelling Breast Cancer is the most frequent disease as a cancer type
typical and powerful non-linear functions. It is consisted of a for women. Therefore, any development for diagnosis and
network of lots of artificial neurons. Each of these prediction of cancer disease is capital important for a healthy
combinations are comprised of input/output characteristics life. In this paper, we have discussed two popular machine
that perform a local mathematical function. The function could learning techniques for Wisconsin Breast Cancer
be a computation of weighted sums of inputs which generates classification. Artificial Neural Network and Support Vector
an output if it goes beyond a given threshold value. The output Machine are used as ML techniques for the classification of
could be an input to other neurons in the network. This WBC (Original) dataset in WEKA tool. The effectiveness of
transaction iterates until the latest output is produced [13]. The applied ML techniques is compared in term of key
authors also published several comparative results in this area performance metrics such as accuracy, precision, recall and
[13, 14]. As machine learning techniques, SVM and ANN are ROC area. Based on the performance metrics of the applied
applied with WEKA machine learning tool. WEKA is Java ML techniques, SVM (Sequential Minimal Optimization
based and open source tool. It provides many machine Algorithm) has showed the best performance in the accuracy
learning algorithms and methods for analysis. It contains many of 96,9957 % for the diagnosis and prediction from WBC
machine learning tools for classification, clustering, dataset.
regression, association rules mining and visualization. In this
study, ARFF (Attribute-Relation File Format) was used for REFERENCES
classification of breast cancer. SMO (Sequential Minimal [1] Cancer, https://www.who.int/en/news-room/fact-sheets/detail/cancer. Last
Optimization) algorithm and LibSVM are used as the Access: 25.01.2019.
classification of SVM in Weka software. Also, Multi-Layer [2] Siegel, R. L., Miller, K. D., & Jemal, A. (2018). Cancer statistics, Ca-a
Cancer Journal for Clinicians, 68 (1), pp. 7-30.
[3] Maity, N. G., & Das, S. (2017). Machine learning for improved diagnosis [10] Huang, M. W., Chen, C. W., Lin, W. C., Ke, S. W., & Tsai, C. F. (2017).
and prognosis in healthcare. In 2017 IEEE Aerospace Conference, pp. 1-9. SVM and SVM ensembles in breast cancer prediction. PloS one, 12 (1).
[4] Asri, H., Mousannif, H., Al Moatassime, H., & Noel, T. (2016). Using [11] Kourou, K., Exarchos, T. P., Exarchos, K. P., Karamouzis, M. V., &
machine learning algorithms for breast cancer risk prediction and Fotiadis, D. I. (2015). Machine learning applications in cancer prognosis and
diagnosis. Procedia Computer Science, 83, pp. 1064-1069. prediction, Computational and structural biotechnology Journal, 13, pp. 8-17.
[5] Bazazeh, D., & Shubair, R. (2016). Comparative study of machine [12] Umadevi, S., & Marseline, K. J. (2017). A survey on data mining
learning algorithms for breast cancer detection and diagnosis. In 2016 5th classification algorithms. In 2017 International Conference on Signal
International Conference on Electronic Devices, Systems and Processing and Communication, pp. 264-268.
Applications, pp. 1-4. [13] Padmapriya S., Devika M., Meena V., Dheebikaa S. B., & Vinodhini R.
[6] Ahmad, L. G., Eshlaghy, A. T., Poorebrahimi, A., Ebrahimi, M., & (2016). Survey on Breast Cancer Detection Using Weka Tool. Imperial
Razavi, A. R. (2013). Using three machine learning techniques for predicting Journal of Interdisciplinary Research (IJIR), Vol 2, no. 4.
breast cancer recurrence. J Health Med Inform, 4 (124). [14] S. Turgut, M. Dagtekin, T. Ensari, Microarray Breast Cancer Data
[7] Bektas, B., & Babur, S. (2016). Machine learning based performance Classification Using Machine Learning Methods, Int. Conf. on Electric
development for diagnosis of breast cancer, Medical Technologies National Electronics, Computer Science, Biomedical Engineerings' Meeting, DOI:
Congress, pp. 1-4. 10.1109/EBBT.2018.8391468, April 18-19, 2018.
[8] Chen, H. L., Yang, B., Liu, J., & Liu, D. Y. (2011). A support vector [15] M. Amrane, S. Oukid, I. Gagaoua, T. Ensari, Breast Cancer
machine classifier with rough set-based feature selection for breast cancer Classification Using Machine Learning, Int. Conf. on Electric Electronics,
diagnosis. Expert Systems with Applications, 38 (7), pp. 9014-9022. Computer Science, Biomedical Engineerings' Meeting, DOI:
[9] UCI Breast Cancer Wisconsin (Original) Dataset, 10.1109/EBBT.2018.8391453, April 18-19, 2018.
https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Origina
l%29. Last Access: 30.01.2019.