International Journal of Trend in Scientific Research and Development (IJTSRD)

Volume 4 Issue 6, September-October 2020 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470

Prediction of Cervical Cancer using Machine

Learning and Deep Learning Algorithms
Kayalvizhi. K. R, N Kanimozhi
Department of Computer Science and Engineering,
G.K.M College of Engineering and Technology, Chennai, Tamil Nadu, India

ABSTRACT How to cite this paper: Kayalvizhi. K. R |

As the development of machine learning and deep learning, more and more N Kanimozhi "Prediction of Cervical
people or organizations use multiple algorithms to analyse large collections of Cancer using Machine Learning and Deep
data to produce meaningful results that help to predict behaviour. And this Learning Algorithms"
kind of technology is increasingly used in medical field to predict some severe Published in
illness in their early stage, for example, cervical cancer. Cervical Cancer is one International Journal
of the main reasons of deaths in countries having a low capita income. It is the of Trend in Scientific
second most common cancer in India in women accounting for 22.86% of all Research and
cancer cases in women. It becomes quite complicated while examining a Development (ijtsrd),
patient on the basis of result obtained from various doctor’s preferred test to ISSN: 2456-6470, IJTSRD33378
determine if the patient is positive with the cancer. There were 96,922 new Volume-4 | Issue-6,
cases of cervical cancer diagnosed in India in 2018. Around the globe, around a October 2020, pp.426-430, URL:
quarter of million people die owing to cervical cancer. Screening and different www.ijtsrd.com/papers/ijtsrd33378.pdf
deterministic tests confuse the available Computed Aided Diagnosis (CAD) to
treat the patient correctly for the cancer. Copyright © 2020 by author(s) and
International Journal of Trend in Scientific
Machine learning and Deep learning algorithms are used in this project and Research and Development Journal. This
determine if the patient has cancer based on the analyses of the risk factors is an Open Access article distributed
available in the dataset. Predicting the presence of cervical cancer can help the under the terms of
diagnosis process to start at an early stage and comparing various models will the Creative
help in finding out the best prediction model for predicting the presence of Commons Attribution
cervical cancer effectively. License (CC BY 4.0)
(http://creativecommons.org/licenses/by
KEYWORDS: Cervical Cancer, Machine learning, Deep learning, Logistic /4.0)
regression, SVM, Decision Tree, Random Forest, Deep Neural networks, Dataset
I. INTRODUCTION
Cervical cancer is one of the deadliest cancers which are Lakh populace to 190-260 for every 1 Lakh population by
threatening women’s health all over the world and it is hard 2025.
to observe any sign in the early stage. The uterine cervix is
the lowest portion of a woman's uterus (womb), connecting Most of the Machine learning algorithms play a vital role in
the uterus with the vagina. Cervical cancer occurs when the providing more appropriate remedy for the affected person.
cells of the cervix grow abnormally and invade other tissues Among all, classification (Supervised) and regression,
and organs of the body. When it spreads, this cancer affects clustering (Unsupervised), Deep Neural Network methods
the deeper tissues of the cervix and may have spread to are widely used in diagnosis.
other parts of the body, most notably the lungs, liver,
bladder, vagina, and rectum. But since cervical cancer is The main idea behind this project is by making the
slow-growing, its progression through precancerous changes utilization of machine learning and Deep Neural Network
provides opportunities for early detection, prevention and algorithms, cervical cancer can be predicted and classify in
treatment. Better means of detection have meant a decline in the stream of data analytics. ML is the key components in big
cervical cancer in most of the countries over the decades. data revolution said by McKinsey Global Institute because it
can effectively find out hidden knowledge and unknown
In 2015, E&Y in association with FICCI Flo has released that patterns in the specified datasets.
19% of women has been suffering with breast cancer, 14%
and 7% of women has been suffering with cervical and Machine learning algorithms are classified into two kinds:
uterus cancers. Especially in India, highest incidences are supervised (class label is known initially) and unsupervised
recorded in Delhi, Kerala and Tamil Nadu due to overweight, (class label is not known initially). Classification (Logistic
tobacco use, low fruit and vegetable intake, alcohol use and regression, Decision Tree, Naïve Bayes, Neural networks,
lack of physical activity. SVM), regression comes under supervised and clustering (K-
Means, K-Medoids, AGNES, DIANA) comes under
Early screening can reduce cancer rates by 1.5-2.5 times, unsupervised learning.
increases five-year survival rate by 3 to 17 times and also
decreases the treatment cost. Likewise, it extends growth In this project, by applying supervised algorithms like
rate among ladies in India to increment from 110 for each 1 Logistic Regression (LR), Decision Tree Classifier (DT),

@ IJTSRD | Unique Paper ID – IJTSRD33378 | Volume – 4 | Issue – 6 | September-October 2020 Page 426
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
Random Forest Classifier (RF), Support Vector Machine
(SVM) and Neural networks, persons with cervical cancer
will be classified. It obviously reduces further proceeding
tests.

II. RELATED WORK

There are many researches with the same topic for the
Prediction of cervical cancer using machine learning and
deep learning algorithms.

Survey 1: WEN WU “Data-Driven Diagnosis of Cervical

Cancer with Support Vector Machine- Based Approaches”,
Department of Blood Transfusion, Jinan Military General
Hospital, Jinan, China Year: 2017 [1]. An advantage of this
paper is Support vector machine (SVM) approach is
A. Data Collection
introduced for cervical cancer diagnosis. Two improved SVM
The dataset, “Cervical Cancer Risk Factors for Biopsy” was
methods, support vector machine-recursive feature
obtained from the UCI Repository. The dataset contains
elimination and support vector machine-principal
habits, demographic information, and medical history of 858
component analysis (SVM-PCA), are proposed to diagnose
patients from the hospital. There are many missing values in
the cancer samples.
this dataset, due to many patients not answering questions
because of privacy concerns. The dataset consists of 858
Survey 2: Yasha Singh, Dhruv Srivastava, P.S. Chandranand &
instances, with 36 attributes. The dataset consists of 36
Dr. Surinder Singh “Algorithms for screening of Cervical
variables and records of 858 women patients. Of the 36
Cancer: A chronological review” [2]. Advantages of this
variables 4 variables are the target variables.
paper are the critical review of different research papers
published that integrated AI methods in screening cervical
No. Attribute Type
cancer via different approaches analysed in terms of typical
metrics like dataset size, drawbacks, accuracy etc. 1 Age Int
2 Number of sexual partners Int
Survey 3: Xiaoyu Deng, Yan Luo, Cong Wang “Analysis of Risk 3 First sexual intercourse Int
Factors for Cervical Cancer Based on Machine Learning 4 Number of pregnancies Int
Methods”, School of Automation, Beijing University of Posts 5 Smokes Bool
and Telecommunications, Beijing, 100876, China [3]. 6 Smokes(years) Bool
Advantages are some risk factors of cervical cancer are 7 Smokes(pack/year) Bool
introduced and three kinds of machine learning algorithms 8 Hormonal Contraceptives Bool
are used to classification of the cervical cancer dataset from 9 Hormonal Contraceptives (years) Int
UCI. 10 IUD Bool
11 IUD (years) Int
Survey 4: Prediction of Cervical Cancer using Voting and 12 STDs Bool
DNN Classifiers, Komala Rayavarapu, Krishna Kishore K.V, 13 STDs(number) Int
Vignan’s Foundation for Science, Technology and 14 STDs: condylomatosis Bool
Research [4]. 15 STDs: cervical condylomatosis Bool
16 STDs: vaginal condylomatosis Bool
Advantages are by applying supervised algorithms like 17 STDs: vulvo-perineal condylomatosis Bool
Logistic Regression (LR), Decision Tree Classifier (DT), Naïve 18 STDs: syphilis Bool
Bayes Classifier (Naïve), Random Forest Classifier (RF), K- 19 STDs: pelvic inflammatory Bool
Nearest Neighbor (KNN), Support Vector Machine (SVM) and 20 STDs: genital herpes Bool
Deep Neural Networks (DNN), persons with cervical cancer 21 STDs: molluscum contagiosum Bool
will be classified.
22 STDs: AIDS Bool
23 STDs: HIV Bool
Drawbacks of these approaches are High Computational cost,
24 STDs: Hepatitis B Bool
Low prediction accuracy, combinations of machine and deep
learning algorithms was not considered. 25 STDs: HPV Bool
26 STDs: Number of diagnosis Int
III. SYSTEM ARCHITECTURE 27 STDs: Time since first diagnosis Int
The process of training a Machine and Deep Learning model 28 STDs: Time since last diagnosis Int
involves providing a Machine and Deep Learning algorithm 29 Dx: Cancer Bool
with training data to learn from. The term Machine Learning 30 Dx: CIN Bool
model refers to the model artefact that is created by the 31 Dx: HPV Bool
training process. 32 Dx Bool

TARGET VARIABLES
The target variables are four tests and are characterized by
‘1’ or ‘0’ in our data set where ‘1’ represents a malignant
tumor and ‘0’ indicates benign tumor.

@ IJTSRD | Unique Paper ID – IJTSRD33378 | Volume – 4 | Issue – 6 | September-October 2020 Page 427
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
They are detailed below: Fig 2- Count of Target Variables Vs Age
SCHILLER
In this test, Schiller's iodine solution is applied to the cervix
under direct vision. Normal cervical mucosa contains
glycogen and stains brown, whereas abnormal or cancer
affected areas do not take up the stain. The abnormal areas
can then be biopsied and examined histologically. The
composition of Schiller's iodine is the same as Lugol's iodine,
but Lugol's iodine being more concentrated. When Schiller's
iodine is not available, Lugol's iodine can be used as an
alternative. Schiller's test is not specific for cervical cancer,
as areas of inflammation and keratosis may also not take up
the stain.

CYTOLOGY
The medical and scientific study of cells. Cytology refers to a
branch of pathology that deals with diagnoses of diseases C. Feature Engineering And Data Visualization
and conditions through the examination of tissue samples During feature engineering process, data is classified based
from the body. on the measurement levels as Numerical data and
Categorical data. Target variables are classified separately.
Cytological examinations are performed on body fluids Numerical data are represented as numbers. Features which
(examples are urine, blood and cerebrospinal fluid) or on cannot be grouped are classified under numerical data.
material that is aspirated (drawn out via suction into a Categorical data describes categories or groups and also
syringe) from the body. Screening is performed using answers to yes or no questions.
cervical cytology (Pap test) or a human papillomavirus
(HPV) test, or a combination of the two tests. Fig 3- Heatmap Correlation: Numerical & Categorical
Data
HINSELMANN
Colposcopic examination was almost impossible to perform
because of the distance from the focus that was no more than
80 mm. Hinselmann test tried to solve this problem by
pulling out the uterine cervix. The examined part is
anemised by this procedure, which can prejudice the final
result and a small amount of blood might also leak. Besides
that, a patient can feel pain if the portio is held by thin
forceps.

BIOPSY
Cervical biopsy is a procedure to remove tissue from the
cervix to test for abnormal or precancerous conditions, or
cervical cancer. Cervical biopsies can be done in several
ways. The biopsy can remove a tissue sample for testing.

B. Exploratory Data Analysis

EDA is a general approach to exploring datasets by means of
simple summary statistics and graphic visualizations in
order to gain a deeper understanding of the data.
D. Re-Sampling Data
Fig 1- Distribution of Target Variables Dealing with imbalanced datasets require strategies such as
improving classification algorithms or balancing classes in
the training data (data pre-processing) before providing the
input data to the machine learning algorithm. The balancing
classes technique is preferred as it has wider application.

The main objective of balancing classes is to either increase

the frequency of the minority class or decrease the frequency
of the majority class. The balancing is done in order to obtain
approximately the same number of instances for both the
classes.

E. Build Models
Modelling was done on the original data after default data
cleaning and scaling where necessary. For all the classifier
algorithms, the dataset is split 25%. 75% for training and
25% for testing. Feature selection is done and the models are
built based on the corresponding features.

@ IJTSRD | Unique Paper ID – IJTSRD33378 | Volume – 4 | Issue – 6 | September-October 2020 Page 428
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
Logistic regression, SVM, Decision tree, Random forests and However, while working in an imbalanced data accuracy is
Deep neural network (DNN) were the algorithms used. not an appropriate measure to evaluate model performance.
To fully evaluate the effectiveness of our model, we must
F. Model Evaluation and Testing examine precision and recall as well.
The conventional evaluation methods do not accurately
measure model performance when faced with imbalanced
datasets.

Standard classifier algorithms like Decision Tree and Logistic

Regression are biased towards classes which have number of IV. RESULTS AND DISCUSSIONS
instances. They can predict only the majority class data. Since we are dealing with medical data, we will be focusing
There is a high probability of wrong classification of the on Recall rather than Precision i.e, we will allow Type I error
minority class as compared to the majority class. to creep in to our models because in our case, Type II error is
far costlier.
Performance evaluation of a classification algorithm is
measured by the Confusion Matrix which contains Being diagnosed with cancer when you don’t have cancer
information about the actual and the predicted class. (false positive), although not desirable, is far better than being
diagnosed that you don’t have cancer when you do (false
Fig 4- Confusion Matrix negative | Type II Error) – Even costing a life.

From the original dataset, we had good precision and recall

but the accuracy is very less for Logistic Regression and SVM
and reasonably good for Decision Tree, Random Forest and
DNN. However the confusion matrix showed that the model
is not able to identify the affected people at a reasonable rate
except DNN.

Algorithm Accuracy (%) Precision 0 Precision1 Recall 0 Recall 1 F1 1 F1 2

Logistic Regression 75% 0.94 0.16 0.78 0.44 0.85 0.244

SVM 82% 0.94 0.21 0.87 0.38 0.90 0.27

Decision Tree 85% 0.93 0.21 0.91 0.25 0.92 0.23

Random Forest 91% 0.92 0.50 0.99 0.06 0.95 0.11

Deep Neural Networks (DNN) 91% 0.91 1.00 1.00 0.13 0.95 0.23

We cannot rely on these and I further go on with resampling V. CONCLUSIONS

the data using SMOTE and above is the result of resampled Cervical Cancer is the cancer arising from the cervix. Usually
dataset. it is very difficult to identify cancer at early stages. The early
stages of cancer are completely free of symptoms. It is only
Even after resampling the data, accuracy is less for Logistic during the later stages of cancer that symptoms appear.
Regression and SVM. Accuracy of Decision Tree has also Predicting the presence of cervical cancer can help the
dropped but we have good precision and recall. diagnosis process to start at an early stage.

We get a reasonable accuracy and recall rate for Random We get a reasonable accuracy and recall rate for Random
Forest Model with both original data and resampled data. Forest Model with both original data and resampled data and
However resampling is not applicable for DNN. DNN.

Confusion Matrix for Random Forest: Hence it’s been concluded that the Random Forest with
Accuracy: 0.912087912879121 SMOTE and DNN are the overall best model so far for
[ [165 1] predicting the cancer indicator if all the 4 target variables
[15 1] ] (Biopsy, Cytology, Hinselmann and Schiller) are combined
together and classified as multi classifier for target variable.
Confusion Matrix for DNN:
[ [192 0] VI. REFERENCES
[ 20 3] ] [1] WEN WU “Data-Driven Diagnosis of Cervical Cancer
with Support Vector Machine- Based Approaches”,
The confusion matrix showed that both Random Forest and Department of Blood Transfusion, Jinan Military
DNN are able to identify the affected people at a reasonable General Hospital, Jinan, China Year: 2017
rate compared to other algorithms.

@ IJTSRD | Unique Paper ID – IJTSRD33378 | Volume – 4 | Issue – 6 | September-October 2020 Page 429
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
[2] Yasha Singh, Dhruv Srivastava, P.S. Chandranand & Dr. [5] Prediction of Cervical Cancer using Voting and DNN
Surinder Singh “Algorithms for screening of Cervical Classifiers Publication: Komala Rayavarapu, Krishna
Cancer: A chronological review” Kishore K.V | Vignan’s Foundation for Science,
Technology and Research
[3] Xiaoyu Deng, Yan Luo, Cong Wang “Analysis of Risk
Factors for Cervical Cancer Based on Machine Learning [6] Dhwaani Parikh, Vineet Menon “Machine Learning
Methods”, School of Automation, Beijing University of Applied to Cervical Cancer Data”, RMIT University, 124
Posts and Telecommunications, Beijing, 100876, China La Trobe St, Melbourne VIC 3000
[4] Upasana “Handle Imbalanced Classification Problems
in machine learning”, Consultant of Data & Analytics in
KPMG

@ IJTSRD | Unique Paper ID – IJTSRD33378 | Volume – 4 | Issue – 6 | September-October 2020 Page 430