Content uploaded by Afeez Adekunle Soladoye
Author content
All content in this area was uploaded by Afeez Adekunle Soladoye on Apr 29, 2024
Content may be subject to copyright.
https://dx.doi.org/10.4314/dujopas.v10i1a.4
ISSN (Print): 2476-8316
ISSN (Online): 2635-3490
Dutse Journal of Pure and Applied Sciences (DUJOPAS), Vol. 10 No. 1a March 2024
*Author for Correspondence
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 29
Detection of Cervical Cancer Using Deep Transfer Learning
Bolaji A. Omodunbi, Afeez A. Soladoye*, Adebimpe O. Esan,
Nnamdi S. Okomba, Temiloluwa G. Olowo and Oluwapelumi M. Ojelabi
Department of Computer Engineering,
Federal University Oye-Ekiti.
Nigeria
Email: sabdulhafeedh@gmail.com
Abstract
Cervical cancer, the fourth most prevalent disease among females, is one of the major threats to the
health of women worldwide. The early detection of cervical cancer is crucial for efficient disease
management and medical care since it increases the likelihood of treating and curing the disease.
Medically cervical cancer is usually diagnosed using pap test or colposcopic examination. This tests are
time consuming and lack of enough, but with technological advancement and application of artificial
intelligence to health care, machine learning algorithms has be giving faster and more accurate
prediction of illness. This study made use of the readily available SIPAKMed dataset which includes
the collection of a large image dataset of 4049 images. Due to the fact that detection of this type of cancer
using medical images gives a faster and more accurate detection compared to using structured dataset
comprising of various attributes related to cervical cancer. Pretrained EfficientNet-B7 was used as the
detection model. The dataset went through series of preprocessing techniques like filtering using low
and high pass filter, normalization using Min-Max Scalar. Label and image data were then converted
to an array. The developed system was evaluated and gave an average accuracy and precision of 87%
and 87% respectively, it was then compared with Graph Neural Network and Random Forest but they
were both outperformed. As a result of this, EfficientNet-B7 is a very good deep learning algorithm for
detecting cervical cancer using medical images. Other architectures of convolutional neural networks
should be experimented and their performance should be compared with the developed system in future
work, while larger dataset should be utilized as well.
Keywords: Convolutional neural networks, EfficientNet B7, Cervical cancer, Medical
imaging, Transfer Learning.
INTRODUCTION
Cervical cancer is the world's fourth most common malignant disease in women. Sub-Saharan
Africa has the highest rates of cervical cancer in the world, largely attributed to low cervical
cancer screening coverage (Pimple and Mishra, 2022). Cervical cancer is a preventable disease.
Yet it is the most common cause of cancer in the African Region where it accounts for 22% of
all female cancers and 12% of all newly diagnosed cancer in both men and women every year
(World Health Organization, 2015). In Africa, 34 out of every 100 000 women are diagnosed
with cervical cancer and 23 out of every 100 000 women die from cervical cancer every year
(Africa Health Organization, 2020). Cancer is a group of diseases involving abnormal cell
growth with the potential to invade or spread to other parts of the body. Cervical cancer is a
type of cancer that occurs in the cells of the cervix – the lower part of the uterus that connects
to the vagina (MayoClinic, 2022). Many factors, such as the human papillomavirus and
sexually transmitted diseases, increase the risk of developing cervical cancer as well as
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 30
smoking. Identifying those factors and developing a classification model to determine
whether the cases are cervical cancer is difficult (Zhang et al., 2020).
Detection of cervical cancer is usually done by physical examination of the cervix, diagnosing
through manual cross examination on Magnetic Resonance Imaging/Computer Tomography
(MRI/CT) scans, Pap test and colposcopic examination. This might be time consuming and
non-availability of enough medical practitioners with specialization in this area might be
challenging. The availability of electronic medical records has made medical prediction easier,
with the use of patient historical data, future occurrence can be predicted ahead of time. As a
result of this, employing machine learning approach to detect cervical cancer in MRI will help
in earlier, faster and more accurate detection of the cancer and in turn help medical doctors to
perform their diagnosis faster and begin treatment plan earlier (Al Mudawi and Alazeb, 2022).
Convolutional neural networks (CNN) have been proven to be a good deep learning
algorithm for image classification due to its good performance on many medical image
classifications. CNN is designed to process data with grid pattern or structured array of data
like images, which has its basis from the formation and arrangement of animal visual cortex
(Soladoye, 2023). It comprises of building blocks like convolution layers, pooling layers, and
fully connected layers.The convolution layer is the basic part of the CNN which is in charge
of feature extraction from the input data, which is the combination of both linear (Convolution
operation) and non-linear operations (activation function). The pooling layer is the layer
where the dimension of the features extracted by the convolution layer is reduced in order to
reduce the size and number of the subsequent parameters that can be learnt (Nisha and
Meeral, 2021). EfficientNet B7 is a CNN based pre trained model on ImageNet-1k, with these
images having a resolution of 600x600. The baseline model of the architecture is B0 designed
by AutoML MNAS, this EffiecientNet-B7 was obtained through the scaling up of the baseline
EffiecientNet-B0 (Khalil et al., 2022).
Section 2 of this research article presents some theoretical background and a review of some
selected and relevant related works. The methodology employed was discussed in Section 3,
while the evaluation result of the system was discussed in section 4. Section 5 concludes the
paper with highlight of the major finding and recommendation for future work.
Cervical cancer typically develops slowly over many years, beginning with pre-cancerous
changes in cervix cells that can be detected through regular cervical cancer screenings, such
as a Pap test or HPV test. These precancerous changes are frequently treatable before they
progress to cancer. If left untreated, cervical cancer can spread to other parts of the body,
including the bladder, rectum, and lungs, and can be fatal. Depending on the stage and
location of the cancer, the main treatments for cervical cancer include surgery, radiation
therapy, and chemotherapy. Early detection and diagnosis of cervical cancer can improve
patient survival significantly (Zhang et al., 2020).
Staging characterises or categorises cancer depending on how much cancer is present in the
body and where it is at the time of diagnosis, as shown in Figure 1, different stages shows the
severity of this cancer based on their stages. Some of the Cervical Cancer signs and symptoms
are abnormal vagina bleeding, unusual vagina discharge, early menopause, narrowing of the
vagina and Lymphoedema (Obermair, 2017).
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 31
Figure 1: Stages of Cervical Cancer (Obermair, 2017)
Many researches have been done on the detection of cervical cancer using various machine
learning techniques. These studies are reviewed to understand the state of art in the research
area and identify possible gaps to fill with the research.
Alsubai, Alqahtani and Sha (2023) proposed a Privacy Preserved Cervical Cancer Detection
Using Convolutional Neural Networks Applied to Pap Smear Images using the publicly
available SIPaKMeD dataset having five cell categories: superficial-intermediate, parabasal,
koilocytotic, metaplastic, and dyskeratotic. Pap smear images were segmented, and a deep
CNN using four convolutional layers was applied to the augmented images of cervical cells
obtained from Pap smear slides. The system gave a good performance when evaluated,
however, the system was not tested on a large dataset of real-world Pap smear images.
Glučina, Lorencin, and Anđelić (2023) conducted a study for Cervical Cancer Diagnostics
Using Machine Learning Algorithms and Class Balancing Techniques. In the research,
publicly available cervical cancer data collected on 859 female patients are used. Each sample
consists of 36 input attributes and four different outputs: Hinselmann, Schiller, cytology, and
biopsy. Due to the significant unbalance of the data set, some data oversamplying techniques
were applied, after which traditional machine learning algorithms were used for the
classification. Al Mudawi and Alazeb (2022) proposed a ML-based approach to predict
cervical cancer with a focus on data pre-processing, predictive model selection, and
computational complexity analysis. Series of pre-processing techniques were employed to
ensure the dataset is in good format. Traditional machine learning algorithms like Random
Forest , decision tree, adaptive boosting, and gradient boosting algorithms and Support vector
machine were used as the classifiers, with the first four algorithms giving best performance
followed by SVM.
Lilhore et al., (2022) proposed a hybrid model for detecting cervical cancer that combines
causal analysis and machine learning techniques. The model utilised a causal Bayesian
network (CBN) to identify the most relevant factors associated with cervical cancer and a
random forest classifier to predict the likelihood of cervical cancer and outperformed other
machine learning models such as support vector machines, k-nearest neighbour, and decision
trees when evaluated. Karuparthi and Abishek (2022) conducted a similar study where they
explored the use of data mining techniques such as the Support Vector Algorithm and
Random Forest Algorithm to predict cervical cancer indications using the biopsy test. It
focuses on the use of the Random Forest Algorithm (RF) to deal with imbalanced medical data
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 32
sets. The CART technique and feature randomness produced a better model than individual
constituent models. Surendiran, Balaji and Deepa (2022) examined the use of Deep Learning
and Machine Learning techniques to classify cervical cytopathology images and discussed
common deep-learning architectures. The study highlights the need for more complex models
to improve accuracy, as most proposed approaches have been applied to the same dataset.
The review provides valuable insights for researchers in the field to build upon existing work.
Chadaga et al. (2022) designed a computer-aided diagnostic method to screen cervical cancer
patients using a custom stacked ensemble machine learning approach. The research
conducted a deep exploratory analysis and utilised techniques for feature selection and
imbalanced data. It also highlights risk factors such as long-term use of hormonal
contraceptives and age and the role of IUDs in decreasing the likelihood of contracting cervical
cancer. . The final model achieved high accuracy, precision, recall, F1-score, AUC, and average
precision. Abdelrahman, Abdelrazek and Eldeib (2021) conducted a performance evaluation
study where he compared the performance of four machine learning algorithms (Logistic
Regression, Decision Tree, Random Forest, and SVM) for predicting cervical cancer using data
from the NICPR. The Random Forest algorithm had the highest accuracy.
Wang et al. (2021) presented a study on predicting cervical cancer using machine learning
techniques. A dataset of cervical cancer patients and non-cancerous controls was used to
develop a model that could predict whether a patient had cervical cancer or not. They used
several feature selection techniques, including the mRMR and CFS algorithms, to identify the
most important features for predicting cervical cancer. These techniques helped to reduce the
number of features needed for accurate prediction and improved the accuracy of the
prediction. Their approach achieved a high accuracy rate for predicting cervical cancer using
the random forest algorithm and improved feature selection technique. Bhatti, Shahzad and
Asif (2021) conducted a similar study to evaluate the performance of the random forest
algorithm in predicting cervical cancer using clinical and demographic data. The authors
collected data from 858 patients and divided it into training (70%) and testing sets (30%). They
applied the random forest algorithm to the data and evaluated its performance using various
metrics. The results showed that the algorithm achieved high accuracy, sensitivity, specificity,
positive and negative predictive values, and area under the receiver operating characteristic
curve. The study concluded that the random forest algorithm could be a useful tool in
predicting cervical cancer, but further research is needed to validate these findings with
additional biomarkers and in larger datasets.
Hussein, Mohamed and Aziz (2021) conducted a predictive study for detection of cervical
cancer, using Random Forest as the machine learning algorithms. Attributes related to
demographic and clinical risk factors were considered as the input phenotypes for the system.
The model was trained on data from 1,255 women in Egypt, including age, education, marital
status, family history, smoking status, and history of sexually transmitted infections. The
evaluated system shows that the Random Forest gave a good prediction performance.
Reviewing all the reported studies, it would be observed that none of the studies was said to
have used Convolutional Neural Networks, as majority of the studies employed traditional
machine learning algorithm. Additionally, most of the used dataset were structured dataset.
This is the reason this study developed a system that employed deep learning algorithm with
a medical image dataset. This will help in proving a broader scope for the state of art of the
study and give an insight of what the performance could be when such technique is employed.
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 33
MATERIALS AND METHOD
The Cervical cancer detection procedure typically consists of four main stages: data
acquisition, data pre-processing and image processing, feature extraction and dimensionality
reduction, model training and evaluation. Data acquisition is concerned with the collection of
isolated cervical Pap smear slides. Pre-processing aims to reduce picture interference and
standardise image size and form to ensure consistent dimensions. This is critical for
improving detection accuracy. During the feature extraction step, important features from the
cervical cancer pictures are extracted. The dimensionality reduction procedure is employed
after the feature extraction step to improve classification accuracy. The algorithm is trained
and evaluated by feeding it labelled vector data that has been flattened. The evaluation phase
focuses on determining how well the model works when supplied with previously unknown
data. Figure 3.1 below illustrates a block diagram outlining the stages involved in the cervical
cancer detection system. These stages include data acquisition, data preprocessing and image
processing feature extraction and dimensionality reduction as model training and evaluation
Figure 1: Overview of the research methodology
Data Acquisition
The dataset used was obtained from Kaggle. SIPaKMeD Database comprises 4049 images of
cells that have been manually extracted from 966 groupings of cells, in Pap smear slides. These
images were captured using a CCD camera attached to a microscope. The cell images are
classified into five categories; superficial intermediate cells, parabasal cells, metaplastic cells,
dyskeratotic cells, and koilocytotic cells including normal, abnormal and benign cells. Table 1
provides detailed information of the SIPaKMeD dataset. Table 1 shows the distribution of this
dataset based on the type of cells captured in it. As classified in Table 1, superficial-
Intermediate cells and parabasal cells are normal cells, metaplastic cells are benign cells while
dyskeratotic cells and koilocytotic cells are abnormal cells.
Image
Acquisition
Image Pre-
processing
Detection
Evaluation
Kaggle,
SIPaKMeD
Database
Scaling
Filtering
Normalization
EfficientNet
Accuracy
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 34
Table 1: Class distribution of different cells in the dataset
Class
Number of Images
Number of Cells
Normal cells
Superficial-Intermediate cells
126
831
Parabasal cells
108
787
Benign Cells
Metaplastic cells
271
793
Abnormal Cells
Dyskeratotic cells
223
813
Koilocytotic cells
238
825
Total
966
4049
Image Pre-processing
The dataset was taken through series of image pre-processing techniques in order to ensure it
was in the right format for processing. The images were firstly resized so that they fell in the
same input size as the model, afterwards, they were filtered using low and high pass filtering.
This was done to remove noise that might lead to distortion of the image. The ‘labels’ and
‘imgdata’ lists were converted into NumPy arrays. Data normalization was then performed
using Min-Max normalization technique.
Detection using EfficientNet
This study employed transfer learning technique, using trained EfficientNet model for
classification of the cervical cancer images. The pre-trained EfficientNet model was fine-tuned
with parameters and features including architecture, optimization, loss function, and batch
size, number of epochs, image dimensions, and filters. The sequential model design
seamlessly integrates layers for feature extraction, combining convolutions with ReLU
activation, max-pooling, and fully connected components (Khalil et al., 2022). Fine-tuning
employed Adam optimizer with a specific learning rate of 0.001, and the sparse categorical
cross-entropy loss function was employed for accurate alignment between predictions and
true labels. Training efficiently occured in batches of 32 images over 200 epochs, achieving a
balance between computational efficiency and accuracy refinement. Images were resized to a
manageable 32 x 32 pixels, finding a pragmatic equilibrium between data quality and
processing demands. The strategic integration of diverse filter sizes in convolutional layers
enhances the model's capacity to detect features of different scales.
Evaluation method and Metrics
The system was evaluated using hold-out evaluation method with 60% of the whole dataset
assigned for the training, 20% for validation and the remaining dataset was used for testing.
After the model was well trained and tested, it was then evaluated using accuracy, precision,
f1score and recall. These evaluation metrics results would be gotten through the printed
classification report after the model have been tested with the test data split. The mathematical
formulation of some of these metrics are represented with Equation 1, 2 and 3.
(1)
(2)
(3)
RESULTS AND DISCUSSION
The developed cervical cancer detection system was developed using EfficientNetB7 as the
detection model. As mentioned in the earlier section. The system was trained with batch size
of 32 over 200 epochs.
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 35
Figure 2 shows the training and validation accuracies plotted against Epoch as the model was
trained over 200 epochs. As seen in the graph, the training accuracy is more than the validation
accuracy with small difference. This shows that our model was well trained and there was
nothing like overfitting of model. This helped in having a good performance of the system
when evaluated. Though the validation accuracy started way too higher than the starting
value of the training accuracy and this made us to know that the training actually considered
the accuracy of the system from the possible values without overlooking the smaller ones.
Figure 2: Training and Validation Accuracies per Epoch
Figure 3 shows the training and validation losses plotted against Epoch as the model was
trained over 200 epochs. It can be seen in the graph that the training loss is below the
validation accuracy with small difference that is less than 20%. This shows that our model was
well trained and there was nothing like underfitting of model. This helped in having a good
performance of the system when evaluated.
Figure 3: Training and Validation losses per Epoch
As earlier mentioned the system was evaluated using some evaluation metric. This system
gave an average accuracy, precision, and F1score of 87, 87 and 87% respectively. Dyskeratotic
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 36
cells class gave the highest precision of 98% and F1score of 92%. The summary of these results
is presented in Table 2 which shows the classification report gotten during system evaluation.
Table 2: Classification Report of the Developed System
Class
Precision
Recall
F1-score
Support
0
0.78
0.86
0.82
160
1
0.85
0.88
0.86
161
2
0.90
0.95
0.92
170
3
0.98
0.89
0.93
152
4
0.85
0.77
0.81
167
Accuracy
0.87
810
macro average
0.87
0.87
0.87
810
Weighted average
0.87
0.87
0.87
810
Comparison with Machine Learning Algorithms
The performance of the developed system trained with EfficientNetB7 was compared with
the performance of Graph Neural Network and Random Forest, so as the showcase the
effectiveness of the developed system. When compared, the developed system outperformed
the other two machine learning algorithms trained with the same dataset. Table 3 shows the
comparison results of the compared algorithms.
Table 3: Experimental results of the Comparison with some Machine Learning Algorithms
S/N
Algorithm
Accuracy (%)
F1score (%)
1
Developed system (EfficientNet B7)
87
87
2
Graph Neural Network
51
51
3
Random Forest
83
83
As shown in Table 3, the developed system is 4% more accurate than random forest and 36%
more accurate than the graph neural network. The 87% accuracy gotten from the developed
system signifies that EfficientNet B7 performed well when used for detection of cervical
cancer with medical image dataset.
CONCLUSION AND RECOMMENDATIONS
This system was developed by employing the flexibility of transfer learning where
EffiecientNet B7 was used for the detection of cervical cancer using the SIPaKMeD Database
that comprises of 4049 images of cells that have been manually extracted from 966 groupings
of cells, in Pap smear slides. The developed system gave an average performance accuracy of
87%, and when compared with some machine learning algorithms like Graph Neural
Network and Random Forest, the two were outperformed by the developed system. This
obviously expresses the efficiency of EfficientNet B7 for the detection of cervical cancer using
medical images. Other architectures of CNN like ResNet, Visual Geometry Group (VGG),
AlexNet and others should be implemented in the future works to check their performance
with the developed system and a larger dataset might be utilized for deep learning of the
model.
REFERENCES
Abdelrahman, M., Abdelrazek, M. and Eldeib, A. (2021). Prediction of cervical cancer using
machine learning algorithms. Informatics in Medicine Unlocked, 23, 100562.
Africa Health Organisation. (2020). Cervical cancer in Africa Fact Sheet. Retrieved August 27,
2023, from https://aho.org/fact-sheets/cervical-cancer-in-africa-fact-sheet/
Al Mudawi, N., &Alazeb, A. (2022). A Model for Predicting Cervical Cancer Using Machine
Learning Algorithms. Healthcare, 10(1), 28: doi: 10.3390/healthcare10010028
Detection of Cervical Cancer Using Deep Transfer Learning
Omodunbi B. A. et al, DUJOPAS 10 (1a): 29-37, 2024 37
Alsubai, S., Alqahtani, A. and Sha. M. (2023). Privacy Preserved Cervical Cancer Detection
Using Convolutional Neural Networks Applied to Pap Smear Images. Advanced
Computational Intelligence System for Secure Medical Imaging Applications.,
https://doi.org/10.1155/2023/9676206
Bhatti, A. I., Shahzad, S. &Asif, M. (2021). Performance analysis of random forest algorithm
for cervical cancer prediction. Journal of Medical Systems, 45(8), 78
Chadaga, A., Wu, R., Tabibian, A. and Chaudhari, A. (2022). Predicting cervical cancer biopsy
results using demographic and epidemiological parameters. Journal of Lower Genital
Tract Disease, 26(1), 1-5.
Glučina, M. Lorencin, A. and Anđelić, N. (2023). Cervical Cancer Diagnostics Using Machine
Learning Algorithms and Class Balancing Techniques. Applied Science,13(2).
https://doi.org/10.3390/app13021061
Hussein, W. M., Mohamed, M. E. and Aziz, M. A. (2021). Development of a random forest
model for predicting cervical cancer risk. Journal of Medical Systems, 45(8), 85.
Karuparthi, S., and Abishek, T. (2022). Comparative Study between Random Forest and
Support Vector Machine Algorithm in Classifying Cervical Cancer. International
Journal of Advanced Science and Technology, 31(4), 7511-7520.
Khalil, M. I., Tehsin, S., Humayun, M., Jhanjhi, N. Z. and AlZain, M. A. (2021). Multi-Scale
Network for Thoracic Organs Segmentation. Computers, Materials & Continua, 70(2),
3251-3265
Lilhore, S., Gupta, D., Singh, R., & Malviya, N. (2022). Hybrid Model for Detection of Cervical
Cancer Using Causal Analysis and Machine Learning Techniques. Journal of Digital
Imaging, 35(1), 119-127. doi: 10.1007/s10278-021-00529-5
Mayo Clinic (2022). Cervical cancer - Symptoms and causes. Retrieved March 11, 2023, from
https://www.mayoclinic.org/diseases-conditions/cervical-cancer/symptoms-
causes/syc-203525
Nisha, S. S. and Meeral, N. M. (2021). Applications of deep learning in biomedical engineering.
Handbook of deep learning in biomedical engineering. https://doi.org/ 10.1016/b978-0-12-823014-
5.00008-9.
Pimple, S. and Mishra, G. (2022). Cancer cervix: Epidemiology and disease burden.
CytoJournal, 19:21
Soladoye, A. A. (2023). Decision Support system for prediction of stroke using Recurrent
Neural Network with Gated Recurrent Units. M.Eng Dissertation, Federal University
Oye-Ekiti. Nigeria.
Surendiran, B., Balaji, G., and Deepa, M. (2022). Exploring the Cervical Cancer Prediction by
Machine Learning and Deep Learning with Artificial Intelligence Approaches. Journal
of Medical Systems, 46(1), 10.
Wang, X., Sun, J., Wu, X. M., Huang, X., Wang, Y. and Wang, Y. (2021). Scalable and Robust
Graph Neural Networks. IEEE Transactions on Neural Networks and Learning Systems,
32(9), 3999-4013.
World Health Organization (2015). Cervical cancer common amongst African women.
Retrived from https://www.afro.who.int/news/cervical-cancer-common-amongst-
african-women
Zhang, S., Xu, H., Zhang, L. and Qiao, Y. (2020) Cervical cancer: Epidemiology, risk factors
and screening. Chinese journal of cancer research,32(6), 720-728
