DEEP LEARNING POWERED CERVICAL DYSPLASIA

CLASSIFICATION: A GRAPH -BASED SEMI SUPERVISED LEARNING

APPROACH WITH YOLOV8 INTEGRATION

Niranjana K Blessy K Dr.M.Selvi

Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
Sathyabama Institute of Science and Sathyabama Institute of Science and Sathyabama Institute of Science and
Technology Technology Technology
Chennai, India Chennai, India Chennai, India
Email: Email:[email protected] Email: [email protected]
[email protected]

ABSTRACT concerned with the utilization of advanced machine

learning and deep learning techniques in developing a
The cervical dysplasia classification project aims to robust system for cervical dysplasia classification,
develop a sophisticated and efficient diagnostic system directed towards assisting healthcare professionals make
by combining state-of-the-art machine learning and accurate and timely diagnoses.
deep learning methodologies. The process starts with
very careful preprocessing of the data, encompassing Cervical dysplasia is the condition characterized by
high-resolution image enhancement, normalization, abnormal cells on the cervix that, if untreated, would
and feature extraction, thus ensuring input quality to eventually become cancerous. Traditional diagnostic tools
the model. The core of the machine learning used to date include Pap smear and colposcopy relying
component is a Graph Convolutional Network (GCN), heavily on human interpretation by skilled pathologists.
which uses graph-based semi-supervised learning to While these are effective methods, they are indeed time-
address issues such as class imbalance and reduce consuming, resource intensive, and prone to error.
reliance on manual annotations. The propagation of Furthermore, in areas where there are fewer numbers of
labels within the GCN framework further enriches the specialized physicians, it means that patients get
training dataset and helps the model to more diagnosed in a later stage. Thus, the system being
accurately classify dysplastic cervical cells. At the proposed here deals with the aforementioned challenges
same time, a retrained model of YOLOv8 is used for by integrating the merits of machine learning and deep
proper feature extraction, utilizing space and learning approaches.
contextual information from an image to enhance the
correctness of classification. The approach is evaluated Firstly, preprocessing of high-resolution cervical cell
with strict metrics such as accuracy, precision, recall, images in order to prepare the data for analysis. The
and cross-validation to ensure its scalability and machine learning component uses a Graph
reliability in real-life applications. By integrating the Convolutional Network (GCN) framework with semi-
strengths of GCN and YOLOv8, this project delivers a supervised learning to minimize manual labeling and
high-accuracy, automated solution for cervical
class imbalance effectively. Techniques of label
dysplasia detection and reducing manual work load .
propagation in the graph framework enhance the dataset
Keywords—Cervical Dysplasia, Graph Convolutional by inferring labels for images which are not labeled.
Network, YOLOv8, Semi-Supervised Learning, Finally, deep retrain in a model based on YOLOv8
Automated Diagnosis. architecture based on the cervical dysplasia dataset to
extract the meaningful spatial and contextual feature
INTRODUCTION
used within the graph-based learning framework for
Cervical cancer remains one of the major causes of death superior classification performances. Integrating both
from cancer among women all over the world, more so in the contextual understanding using the the YOLOv8
developing countries, which have less access to medical system allows an achievable system of high accuracy
facilities. Early detection coupled with timely intervention despite reduction in manual workload coupled with
is, therefore, important in preventing the progression of
dependency on large extents of labeled datasets. The
cervical dysplasia into invasive cancer. This project is
biggest challenge in the project is that the data becomes
complex in the form of preprocessing, handling class

1
 imbalance, and ensuring robustness in feature extraction. The graph-based framework captures the spatial and
structural relationships between images. Each image will
Preprocessing, which includes normalization,
be represented as a node and edges between nodes will
enhancement, and the like, is one of the important represent similarities based on pixel patterns, texture
aspects of the data as it improves the quality of inputs features, and morphological characteristics. Semi-
available to the models. Such techniques as label supervised learning techniques are used to propagate the
propagation and over-sampling balance the set, labels from the labelled nodes to the unlabelled nodes,
addressing inherent bias with majority classe taking advantage of the natural similarities among
images. This approach effectively expands the training
The final goal of this work is to provide a feasible and dataset without requiring extensive manual effort,
automated system that will allow healthcare providers in making it highly scalable and cost-effective.
resource-scarce settings to get an earlier and more The core component of the system is the YOLOv8 model,
accurate cervical dysplasia diagnosis. This system can which is known for its outstanding feature extraction
reduce the burden of cervical cancer by allowing for abilities and speed. Initially pre-trained on large datasets,
timely diagnosis and intervention. This system's YOLOv8 is fine-tuned using the cervical dysplasia
performance is estimated using metrics such as accuracy, dataset to specialize in detecting patterns indicative of
precision, recall, and F1-score, and thus it is reliable dysplasia. It extracts complex features including the size
of cells, texture irregularities, and patterns of structure
enough to be applied in actual scenarios. that help make exact distinctions between normal cells
and abnormal cells. In short, YOLOv8 combined with the
In conclusion, this project integrates advanced graph framework includes the advantages of deep feature
techniques like GCN and YOLOv8 to make a extraction together with spatial relationship models
comprehensive system for the classification of cervical resulting in superior classification accuracies.
dysplasia. It addresses the complexities of data and
dependency on manual labeling, providing a scalable and The classification process gets much stronger support
from a classifier-based neural network, where all the
efficient tool to aid clinicians, thus improving patient
integrated features taken into account from YOLOv8 and
outcomes and reducing global healthcare disparities with the graph-based framework are processed through that
regard to cervical cancer. neural network classifier. Multiple rounds of training,
hyperparameter optimization, and cross-validation
DESCRIPTION techniques such as k-fold validation are used to maximize
performance metrics. Key metrics used for evaluating
This project aims to transform cervical dysplasia robustness and reliability include accuracy, precision,
classification through integration with state-of-the- art recall, and F1-score. These evaluations ensure that the
machine learning approaches, which include graph-based model generalizes well on unseen data, reducing
semi-supervised learning and YOLOv8, which is one of overfitting and increasing practical usability.
the state-of-the-art object detection algorithms. The aim
here is to provide greater accuracy in diagnosis while In addition to the technical aspect, the system emphasizes
highly minimizing manual labelling requirements. usability and accessibility through a Python-based user
Through this system, it addresses many challenges in interface. This user interface is designed to make it easy for
medical imaging datasets that arise from variations in clinicians and medical professionals to upload images,
quality of images, lighting conditions, viewing angles, and visualize preprocessing results, and analyses
morphological complexities. The system will allow for a classification outputs. It also supports interactive
streamlined diagnostic process, reduced human errors, and visualizations that highlight abnormalities, offering
timely results, therefore, making it highly applicable in detailed insights for informed decision-making. The
clinical settings. intuitive design ensures that even users with limited
technical expertise can operate the system effectively, thus
The base of the project comprises collection of high- promoting its adoption in medical environments.
resolution cervical cell images from various sources,
including medical image repositories and clinical Scalability and adaptability are at the heart of the proposed
databases. A dataset collected is given an intense system architecture. Modular design allows for future
preprocessing to ensure its quality and consistency. expansion, including adding more datasets, updated
Preprocessing includes normalization to standardize pixel algorithms, and enhanced functionalities. This makes the
intensity, noise reduction techniques to remove system highly flexible and capable of evolving with
distortions, and data augmentation to artificially increase advancements in AI and medical imaging technologies.
the dataset and enhance its variability. This will make the Moreover, the system can be extended to other diagnostic
model robust to different imaging conditions. In addition, applications, such as classification of various types of
manual and semi-automated labelling methods are used to cancers and abnormalities, making it have a much wider
create a subset of annotated images that act as a training impact in the medical diagnostics domain.
base for supervised learning algorithms.
Beyond its technical capabilities, this project also

2
 emphasizes clinical impact. By enabling early feature selection with generative adversarial networks,
detection and classification of cervical dysplasia, it generating synthetic data for overcoming class
supports timely interventions and better treatment imbalance that increases classification accuracy. Despite
planning. This reduces the likelihood of disease this efficiency, GANs suffer from the computationally
progression, ultimately improving patient outcomes heavy and time- consuming nature, which presents
and reducing the burden on healthcare systems. The difficulties for the real-time clinical applications.
system’s ability to deliver real- time results ensures
that clinicians can make faster and more accurate [3] Upadhyay et al. (2024) evaluated CNN architectures
decisions, improving efficiency in diagnostic such as VGGNet, GoogleNet, and DenseNet121 for
workflows. cervical dysplasia detection. DenseNet121 outperformed
others in accuracy and robustness due to its densely
The system further encompasses the incorporation of advanced connected layers that facilitated feature reuse. However,
security that protects patient data and meets requirements as issues such as overfitting and data scarcity persisted,
stipulated in the health care regulations. For this reason, since a highlighting the importance of improved generalization
medical image deals with sensitive patient data, the system techniques.
adopts encryption techniques for securing information and
storing securely. The system is thus compliant with international [4] Chaudhuri et al. (2023) used label propagation for
standards such as HIPAA and GDPR on proper ethical handling inferring labels of unlabeled data in graph-based semi-
of patient information. supervised learning. The method utilized relational
patterns in graph structures to enhance classification
Another major advantage of this project is its reduction of the performance in cases with few labelled samples.
workload of doctors. It automates the analysis and classification However, the work pointed out that the method is prone
of images, allowing doctors and pathologists to focus on critical to error propagation if the initial labels are not of high
cases. Thus, it saves time and at the same time increases quality.
efficiency in healthcare systems. Besides that, the semi-
supervised learning approach reduces dependency on manually [5] Mahalakshmi et al. (2021) demonstrated the potential
labelled data, thus quickly deploying and adapting to new of transfer learning in reducing computational costs and
datasets without extensive retraining. improving classification accuracy for cervical dysplasia
detection. Transfer learning reduced the need for training
Moreover, the project is designed to scale and be accessible and increased accuracy using pre-trained models. The
using cloud infrastructures. Cloud deployment also allows study highlighted the importance of source datasets
remote access, giving practitioners a chance to work with the closely aligned with target datasets for optimal
system from different locations especially in resource-limited performance.
areas where some medical expertise may be difficult to access.
The interaction with cloud services also allows it to have [6] The authors Reddy et al. (2023) successfully applied
continuous updates and enhancements without disturbing the
Graph Convolutional Networks (GCNs) to represent
ongoing processes.
spatial relationships within cervical dysplasia detection.
GCNs outperformed in a semi-supervised task to
This project provides a base for further research in AI-
preserve spatial information, hence crucial for
based healthcare solutions. It will demonstrate how
classification accuracy; however, their computational
machine learning and neural networks can be used
complexity became a limitation for real-time
together to address the complexities of medicine, thereby
applicability, and therefore this calls for more efficient
opening avenues for future innovation. This framework
architecture.
can be built upon by researchers in new techniques to
handle imbalanced datasets, improve model
[7] Kumar et al. (2020) employed data augmentation
generalizability, and develop hybrid architectures that
techniques, including geometric transformations and
merge multiple AI models.
synthetic data generation, to address data scarcity
LITERATURE SURVEY challenges. These methods enhanced model robustness
and generalization by diversifying training datasets. The
[1] Moldovan in 2020 proposed SVM model with study highlighted the necessity of tailoring augmentation
feature selection to classify small datasets in techniques to the dataset’s characteristics for optimal
cervical dysplasia. The key emphasis was on the results
choice of relevant features, which improved the
model's prediction accuracy while minimizing [8] Patel et al. (2023) highlighted the role of XAI
computational complexity. However, the study techniques, such as SHAP and LIME, in improving
pointed out limitations of SVM in handling high- model transparency and trust. These methods provided
dimensional data and emphasized the requirement insights into decision-making processes, which
for more scalable methods for real-world increased clinician trust and adoption of AI tools in
applications. healthcare, underlining the importance of interpretability
in medical applications.
[2] Elakkiya et al. introduced the hybrid FSOD-
GAN model in 2022. The method integrated
3
 [9] Rajesh et al. (2022) suggested a hybrid model Feature extraction is done using a pre-trained YOLOv8
that combines CNNs with graph-based learning to model, which is particularly known for its accuracy in
enhance the precision of classification. This object detection tasks. In this case, the model is fine-tuned
method reduced the manual requirement for specifically to cervical dysplasia classification, so its
labelling and increased scalability. However, it parameters would be optimized for this task. Fine-tuning
was challenging to integrate such techniques and simply means training the model with the cervical dataset,
required careful parameter tuning to achieve with its weights and biases adjusted such that it can identify
optimal results in a clinical setting. detailed patterns and structures in cervical cells. Further
features like shape, texture, and intensity variations are
[10] Singh et al. (2021) demonstrated the capability of added to the graph-based learning framework to improve
YOLOv8 in object detection and segmentation of the representation of each image. Such features are very
cervical cell images. As it is fast and highly accurate, crucial for differentiating between normal and dysplastic
YOLOv8 showed good potential for use in clinical cells, and thus the model can distinguish subtle variations
practice. Among these, large annotated datasets that that may be indicative of abnormalities. The fine-grained
depend on such systems motivated semi-supervised details captured by YOLOv8 make this variant especially
techniques to supplement training data. appropriate for the case of medical image analysis in which
precision and sensitivity have prime importance. The
Even with all this, cervical dysplasia classification fine-tuning process will tune the model to be better able to
remains a very challenging problem. It still lacks the distinguish between normal and dysplastic cells, paying
potential to be translated into real-time applications in a special attention to key visual cues like cell size, shape, and
clinical environment. texture irregularities.
The training stage of the classifier takes in the features
extracted both by YOLOv8 and the graph-based approach.
PROPOSED SYSTEM A neural network is trained to learn the patterns, which are
The proposed system for cervical dysplasia to differentiate between normal cells and dysplastic ones.
classification is focused on the critical Here, the model is getting relationships and associations
need for the correct and efficient detection of cervic among the features, which enables the system to make
al abnormalities. It includes advanced machine accurate predictions. The training process will contain
learning techniques: graph-based semi-supervised regularization techniques to prevent overfitting and
learning, and the object detection model YOLOv8, enhance the performance. The validation will be done on
to bring high performance with minimal intensive separate datasets; accuracy, precision, recall, and F1-score
manual labelling. By using these approaches, the metrics are computed to analyze how effective the system
system allows for full utilization of both labelled is. These evaluation tests indicate the generalization
and unlabelled data for a scalable and robust capability of the model to unseen data with high
solution suitable for clinical applications. The classification accuracy. Cross-validation techniques
method is based on several interacting stages that further guarantee that the performance of the model is
ensure smooth data processing, feature extraction, stable across different subsets of data, thereby reiterating
and classification. This approach relies on state-of- its reliability for clinical deployment.
art artificial intelligence techniques to achieve
The system proposed in this paper consists of several
optimal detection accuracy, reduce computational
modules working together to achieve the desired result.
complexity, and adapt to various imaging quality The data collection and preprocessing module ensures that
and sample diversities. the data is consistent and of good quality, thereby preparing
The process will begin with high-resolution cervical it for further processing. The graph construction and label
cell images acquired from standard medical datasets. propagation module builds the relationship between GPS
Preprocessing methods are applied to enhance labelled and unlabelled data, which allows for semi-
quality and standardize the various data formats. supervised learning. The YOLOv8 fine-tuning module
These are normalization, augmentation, and adapts the model for cervical dysplasia images, optimizing
denoising. Normalization standardizes the pixel feature extraction. The feature integration module
intensity values, so that consistency is maintained in combines the outputs from graph-based learning and
the dataset. Augmentation introduces variations YOLOv8 to create a unified feature set. The classification
such as rotations, scaling, and flipping, which module trains and tests the neural network model to give
improve the model's ability to generalize. Denoising the predictions based on the features integrated. The
removes background noise, retains key structural deployment module assists in integration with clinical
features in images, and is essential to proper workflows that allow for real-time usage and support
classification. This step of preprocessing forms a performance evaluation of metrics in the model, ensuring
base for further analysis and model building because robustness. A more efficient approach to incorporate
it refines input data, making it even closer to the real labelled and unlabelled data is then possible via the use of
environment of clinical practice. This is an graph structures, which leads to lowering manual
important step for preprocessing because poor- annotation needs and a more feasible semi-supervised
quality input may degrade the performance of a approach towards datasets limited by the numbers of
model. labelled samples.
4
BLOCK DIAGRAM iterative improvements, so the system can adapt to
changing requirements. Stress testing under varying
conditions further tests the robustness of the model so that
it can practically perform in real scenarios.

The deployment phase focuses on the integration of the

system into clinical workflows. The interface, developed in
Python, is user-friendly and facilitates the uploading of
cervical images along with visualizing the preprocessing
steps. The interface offers real-time analysis capabilities
such that predictions are quick and accurate. It also offers
scalability, which means it can be extended to add more
data or future upgrades. This integration will ensure the
proposed system is practical as well as impactful in real
scenarios. The interface also has features for the
visualization of predictions, showing heatmaps, and giving
explanations for classifications, making it more transparent
Fig 1.1 SYSTEM ARCHITECTURE and user-friendly.

The deployment module assists in integration with In conclusion, the proposed system here integrates graph-
clinical workflows that allow for real-time usage and based semi-supervised learning with YOLOv8 to improve
support performance evaluation of metrics in the cervical dysplasia classification. It promotes better feature
model, ensuring robustness. A more efficient extraction, saves efforts in manual labelling, and ensures
approach to incorporate labelled and unlabelled data high accuracy for classification. With its strong
is then possible via the use of graph structures, which preprocessing, scalability, and real-time analysis
leads to lowering manual annotation needs and a capabilities, the system is suitable for clinical workflows,
more feasible semi-supervised approach towards providing reliable and efficient diagnostic support.
datasets limited by the numbers of labelled samples.
Indeed, this can often happen in medical
applications. The better feature extraction ability of METHODOLOGY
the YOLOv8 model ensures high precision and
recall, which are used by the system in order to The methodology of the cervical dysplasia classification
capture subtle variations or abnormalities in cervical project leverages advanced machine learning techniques by
cell structures. It can handle imbalanced datasets, combining graph-based semi-supervised learning and
where abnormal samples are often fewer than YOLOv8 fine-tuning to enhance cervical dysplasia
normal ones, further reinforcing the utility of the classification. It effectively addresses data limitations
system in clinical contexts. Scalability is another through semi-supervised learning and improves detection
strength of this system. The system is designed to accuracy, offering a scalable and reliable solution for
handle large datasets without compromising clinical applications. The system's integration of advanced
performance, thus making it suitable for deployment technologies ensures its potential for broader applications
in medical institutions with high data volumes. Its in medical image analysis.
robustness is due to the advanced preprocessing
1. Data Collection
techniques and optimized feature extraction methods
that enhance performance under diverse conditions The process commences with the gathering of high-
such as lighting, angles, and image quality. The real- resolution images of cervical cells from publicly available
time analysis capabilities enhance its practicality, repositories of medical images and clinical datasets. These
giving the pathologists instant results and the ability images pass through a pre-screening process to ensure
to make timely decisions. This real-time feedback is quality and relevance for the classification of cervical
critical in the early diagnosis and planning of dysplasia. All attention is paid to diversity of the dataset to
treatment, which could drastically improve patient minimize possible biases in the dataset and thus completely
outcomes. represent different stages of dysplasia.
The evaluation phase of the proposed system is very
2. Preprocessing
rigorous testing to validate its performance. Metrics
such as accuracy, precision, recall, and F1-score are Images obtained are preprocessed for uniformity and
used to provide a comprehensive assessment of enhancement quality for analysis. Normalization is applied
classification quality. The confusion matrix visually to scale image intensities uniformly, minimizing variations
represents classification performance, highlighting caused by lighting conditions. Augmentation techniques,
correct predictions, false positives, and false including rotation, flipping, and zooming are utilized to
negatives. This analysis ensures that the system is artificially expand the dataset and introduce variability,
reliable and ready for clinical deployment. helping to improve model robustness. Denoising filters are
Continuous evaluation and feedback loops allow also employed to remove noise and preserve cellular
5
structure clarity, ensuring high-quality input data for cells. The classifier is evaluated using performance metrics
training the model. such as accuracy, precision, recall, and F1-score to ensure
reliability. Validation is performed on an independent test
3. Annotation dataset to assess generalizability and robustness under real-
world conditions.
A subset of the preprocessed images is manually
labeled with known categories of dysplasia-normal, 9. Performance Assessment
low-grade and high-grade dysplasia. Semi-
automated annotation tools have been integrated to The performance of the proposed system is rigorously
speed up the process with guaranteed high labeling assessed through multiple techniques. Cross-validation is
accuracy. It is a hybrid annotation type where the conducted to test model consistency across different data
model can function using a smaller labeled dataset splits, while confusion matrices provide insights into
combined with a larger unlabeled dataset. misclassification rates. ROC curves are analyzed to
determine the trade-off between sensitivity and specificity,
4. Graph Construction to ensure that the model performs well in cervical dysplasia
detection.
A graph-based structure captures relationships
between images. A node represents each image and 10. Deployment
edges connect nodes based on similarity in features
or pixel-level similarity. This structure captures The final model is deployed through a user-friendly
spatial and contextual relationships between interface suitable for clinical use. The interface provides
samples, thus allowing for label propagation in the the clinician with an ability to upload cervical images,
graph. The graph forms a basis for incorporating perform real-time analysis, and receive immediate
both supervised and semi-supervised learning classification results. A report generation module is also
methods for improving classification performance. included to provide visual and statistical feedback for
further diagnosis and decision-making. The system is built
5. Label Propagation for scalability and seamless integration into clinical
workflows.
The graph-based model allows for semi-supervised
learning by propagating labels from labeled nodes to 11. Ethical Considerations
the unlabeled ones. The relationships between
labeled and unlabeled images are used to predict Ethics are at the heart of this approach. Patient data is
categories for the latter. This is done by utilizing the anonymized and encrypted to ensure privacy and security.
graph structure to diffuse information. This helps in Informed consent is obtained for collection and use of data,
reducing dependence on manual annotation and with steps to minimize biases via diversity in the training
enhances the precision of classification by learning datasets. This approach is within the ethical framework and
patterns within the dataset. best practices for medical research and AI applications.
6. YOLOv8 Fine-Tuning
CONCLUSION
The YOLOv8 model, pre-trained on general
datasets, is fine-tuned using the cervical dysplasia In conclusion, the proposed system that is built for cervical
dataset. The pre-trained model serves as a base for dysplasia classification leverages graph-based semi-
feature extraction, and fine-tuning optimizes it to supervised learning and YOLOv8 object detection that
capture domain-specific features of cervical cells. results efficient in both originality and fine accuracy and
This stage improves the ability of the model to that requires little manual labeling. Normalization,
identify patterns and structures in cervical images, augmentation, and denoising are applied to cervical cell
leading to enhanced classification performance. images of high resolution to improve data quality. Its
architecture can be divided into pre-processing, feature
7. Feature Integration extraction and classification that can be done in real time
The fine-tuned YOLOv8 model provides features through a Python API. This capability makes it to handle
that are integrated with graph-based features in order large data set and is resistant to variations in lighting and
to form a representation of each image. It is a hybrid image quality. First, accuracy, precision, and recall ensure
approach because it brings together the precision of its effectiveness in assessments and stress testing
YOLOv8's object detection with the contextual guarantees its dependability. This system improves the
learning of graph-based methods. Both aspects help general classification, real-time analysis and differentiation
improve the classification framework as they capture and makes a more impressive basis for the progression
the localized and relational information within the within the medical field diagnostic imaging and analysis.
dataset.

8. Classification
The integrated features will be used to train a neural
network classifier, to classify normal and dysplastic
6
 Method. International Conference on e-Health
REFERENCES and Bioengineering (EHB). DOI:
10.1109/EHB50910.2020.9280215.
[1] Bhuta, N., Jadhav, T., Shinde, S., Gaikwad,
A., & Jangale, P. (2023). Feature Level [9] Partanen, M., et al. (2018). NordScreen - An
Ensemble Learning Technique for Cervical Interactive Tool for Presenting Cervical Cancer
Cancer Cell Classification. IEEE Screening Indicators in the Nordic Countries.
Conference on Computing, IEEE 14th International Conference on e-Science.
Communication, Control and Automation DOI: 10.1109/eScience.2018.00054.
(ICCUBEA). DOI:
10.1109/ICCUBEA58933.2023.10392076. [10] Shetty, A., & Shah, V. (2018). Survey of Cervical
Cancer Prediction Using Machine Learning: A
[2] Chaudhuri, S., Suneetha, K., & Yadav, R. Comparative Approach. International Conference
K. (2023). Ensemble Learning for Multi- on Computing, Communication, and Networking
Class Classification in Cervical Cancer Technologies (ICCCNT). DOI:
Risk Factor Analysis. IEEE Conference on 10.1109/ICCCNT.2018.8494169.
ICT in Business Industry & Government
(ICTBIG). DOI: [11] Upadhyay, D., Manwal, M., Kukreja, V., &
10.1109/ICTBIG59752.2023.10456058. Sharma, R. (2024). Deep Learning-Based
VGGNet, GoogleNet, and DenseNet121 Models
[3] Djong, A. M., & Widjaja, M. (2023). A for Cervical Cancer Prediction. International
Web-Based Expert System for Diagnosing Conference for Emerging Technology (INCET).
Cervical Cancer Using the Dempster- DOI: 10.1109/INCET61516.2024.1059326
Shafer Method. International Conference
on New Media Studies (CONMEDIA).
DOI:
10.1109/CONMEDIA60526.2023.104281
93.

[4] Elakkiya, R., Subramaniyaswamy, V.,

Vijayakumar, V., & Mahanti, A. (2022).
Cervical Cancer Diagnostics Healthcare
System Using Hybrid Object Detection
Adversarial Networks. IEEE Journal of
Biomedical and Health Informatics. DOI:
10.1109/JBHI.2021.3094311.

[5] Ganguly, T., Pati, P. B., Deepa, K., Singh,

T., & Özer, T. (2023). Machine Learning-
Based Comparative Analysis of Cervical
Cancer Risk Classification Algorithms.
Advances in Computing, Communication,
and Applied Informatics (ACCAI). DOI:
10.1109/ACCAI58221.2023.10200617.

[6] Indra, Z., Jusman, Y., Salambue, R., &

Mahdiyah, E. (2023). Comparison of
Popular CNN Architectures for Cervical
Cancer Early Diagnosis. Artificial
Intelligence and Image Processing
(IWAIIP). DOI:
10.1109/IWAIIP58158.2023.10462887.

[7] Kwatra, C. V., & Kaur, H. (2023). Relative

Analysis of Cutting-Edge Deep Learning
Models for Cervical Cancer Detection.
International Conference on Image
Information Processing (ICIIP). DOI:
10.1109/ICIIP61524.2023.10537644.

[8] Moldovan, D. (2020). Cervical Cancer

Diagnosis Using a Chicken Swarm
Optimization Based Machine Learning
7