Department of Artificial Intelligence and Machine Learning

Capstone Project Part II

Project Title:
“Deep Multi-Model Feature Extraction for
Cervical Cell Classification and Cancer Detection”

By: Guided By:

160121729050: Pranav Raj Bulus Ms. Prachi Dusa
160121729061: V. Bhanu Sai Assistant Professor, Dept. of AIML

1
 LIST OF CONTENTS

● Abstract
● Problem Statement
● Literature Survey
● Completed Objectives
● Architecture
● Modules
● Results
● Conclusion
● Contributions
● References
 ABSTRACT

● CerviDetect is a comprehensive web application that automates cervical cancer detection by analyzing
medical images, such as Pap smear or colposcopy images, using deep learning models like MobileNet,
EfficientNet and ViT for feature extraction and optimization techniques like Genetic Algorithm.
● Built with Python frameworks and libraries, including TensorFlow and OpenCV, the application offers
an interactive interface and is trained on a labeled dataset of cervical cell images, enabling high
accuracy in classification.
● The model effectively identifies normal, pre-cancerous, and cancerous cells, with potential applications
in healthcare for early diagnosis, contributing to better patient outcomes through advanced image
analysis and feature optimization techniques.

3
 PROBLEM STATEMENT

Cervical cancer is one of the leading causes of death among women, and early detection is crucial for
improving survival rates. Current diagnostic methods rely heavily on manual inspection, which can be time-
consuming and prone to human error. There is a need for an automated system that accurately detects cervical
cancer from medical images using advanced deep learning techniques to assist healthcare professionals in
making faster, more reliable diagnoses.

4
Literature Survey

5
Year Author Name Title Journal Methodology Strengths Limitations
Name

2023 S. Vachmanus, T. DeepMetaForge: A IEEE DeepMetaForge The framework Its effectiveness may
Noraset, W. Deep Vision- Access employs BEiT, a achieves an 87.1% be limited when
Piyanonpong, T. Transformer vision transformer, macro-average F1 metadata is
Rattananukrom and Metadata-Fusion as the image- score across four incomplete or
S. Tuarob Network for encoding backbone datasets and is unavailable, and it
Automatic Skin and merges encoded scalable for requires further
Lesion Classification metadata with visual various machine- validation in diverse
features for skin learning clinical settings to
cancer detection. It paradigms, confirm its
simulates how including low- generalizability.
photos with resource devices,
metadata are making it suitable
interpreted, for telemedicine.
improving
diagnostic accuracy.

6
2024 Maram Fahaad An Edge Computing-Based IEEE Access Edge computing, Enhanced accuracy in Limited by edge
Almufareh Factor-Aware Novel Customized VGG16 melanoma detection, real- device computational
Framework for Early architecture, Real- time processing, and capacity, data
Detection and Classification time analysis, Privacy improved privacy dependency, and
of Melanoma Disease preservation preservation hardware constraints
Through a Customized
VGG16 Architecture With
Privacy Preservation and
Real-Time Analysis

2024 Subhajit Early Detection of Multiclass IEEE Access Transfer learning, High accuracy (98%), High computational
Chatterjee, Skin Lesions Using Transfer IncepX-Ensemble improved robustness complexity, reliance
Joon-Min Gil, Learning-Based IncepX- model combining through data augmentation, on data quality,
Yung-Cheol Ensemble Model InceptionV3 and superior performance using limited to the
Byun Xception, Data ensemble learning HAM10000 dataset
augmentation

7
2024 V. S. S. Bala Tripura Pipelined Structure in IEEE AlexNet CNN and High accuracy (98.66% for HAM- Time complexity
Sathvika, Nagilla the Classification of Access SVM model with 10000 dataset), robust segmentation increases with the
Anmisha, Vada Skin Lesions Based on bisectional texture and feature extraction, comparative number of features
Thanmayi, M. AlexNet CNN and feature extraction, analysis of shallow and deep learning (Ns), performance
Suchetha, D. Edwin SVM Model With Bi- segmentation using techniques reliant on optimal
Dhas, S. Sehastrajit, Sectional Texture k-means clustering, feature selection
Sathyanarayanan N. Features classification using parameters (Ns, β),
Aakur the HAM10000 and requires high
PAD-UFES-20 computation for large-
datasets scale datasets

2024 Xin Qian, Tengfei SPCB-Net: A Multi- IEEE ResNet101 and Achieves 97.10% and 99.87% Lacks analysis on
Weng, Qi Han, Chen Scale Skin Cancer Access VGG19 accuracy on HAM10000 and NCT- other classification
Wu, Hongxiang Xu, Image Identification architectures, CRC-HE-100K datasets respectively, tasks, performance
Mingyang Hou, Network Using Self- applied on innovative feature extraction methods reliant on the
Zicheng Qiu, Interactive Attention HAM10000 and (SAP and bilinear-trilinear pooling), computational cost of
Baoping Zhou, Pyramid and Cross- NCT-CRC-HE-100K improves classification accuracy the high-order
Xianqiang Gao Layer Bilinear- datasets compared to state-of-the-art models interaction methods
Trilinear Pooling used in pooling

8
2024 Yutong Xie , A Mutual IEEE TRANSACTIONS The MB-DCNN model The MB-DCNN The MB-DCNN relies
Jianpeng Bootstrapping Model ON MEDICAL IMAGING uses coarse improves segmentation heavily on high-quality
Zhang , Yong for Automated Skin segmentation for and classification via annotated data for both
Xia , and Lesion Segmentation masks, classification, mutual knowledge segmentation and
Chunhua Shen and Classification. and enhanced transfer, effectively classification.
segmentation to handling class and
improve results. pixel imbalance for
better performance.

2024 N. Nigar, M. A Deep Learning IEEE Access The proposed XAI- The model achieves Despite the high
Umar, M. K. Approach Based on based skin lesion high classification performance, the
Shahzad, S. Explainable Artificial classification system accuracy, precision, reliance on LIME for
Islam and D. Intelligence for Skin utilizes deep learning recall, and F1 score, explanations may not
Abalo Lesion Classification to classify eight types making it reliable for fully capture complex
of skin lesions. It early skin cancer decision-making
integrates the LIME diagnosis. The processes.
framework to provide inclusion of Additionally, the
visual explanations for explainability through model's effectiveness
model predictions, LIME boosts trust is tested on a specific
enhancing among dermatologists, dataset, which may
interpretability. aiding in clinical limit generalizability to
decision-making. other clinical
environments.

9
2023 M. Saeed, A. The Power of Generative IEEE Access This study uses CNN The hybrid The reliance on
Naseer, H. AI to Augment for variants to classify skin VGG19+SVM model complex generative
Masood, S. U. Enhanced Skin Cancer lesions, leveraging achieves 96% accuracy techniques for data
Rehman and V. Classification: A Deep GANs and ESRGAN on ISIC 2019, preprocessing may
Gruhn Learning Approach for data augmentation. showcasing the increase computational
Models like VGG16, effectiveness of demands. Additionally,
VGG19, SVM, and combining generative AI while effective on the
hybrid VGG19+SVM with CNN-based transfer ISIC datasets, the
are trained and tested learning for skin cancer model’s performance in
on the ISIC 2019 and classification. real-world clinical
2020 datasets. settings needs further
validation.

2023 A. Magdy, H. Performance IEEE Access This study presents two The proposed methods The use of a small
Hussein, R. F. Enhancement of Skin methods for classifying achieved over 99% dataset of 4,000 images
Abdel-Kader and Cancer Classification dermoscopic images: accuracy in some may limit result
K. A. E. Salam Using Computer Vision KNN with pre-trained models, demonstrating generalizability, and the
deep networks as their effectiveness in model complexity,
feature extractors, and skin cancer detection and particularly with the
AlexNet optimized by a providing a grey wolf optimizer for
grey wolf optimizer. comprehensive AlexNet, raises
evaluation of ML and computational demands.
DL techniques.

10
2023 H. L. Gururaj, N. DeepSkin: A Deep IEEE Access This study uses CNNs excel in object The reliance on the
Manju, A. Learning Approach for CNNs for early skin detection and HAM10000 dataset,
Nagarjun, V. N. Skin Cancer cancer detection with classification, with while comprehensive,
M. Aradhya and Classification the HAM10000 transfer learning using may limit the model's
F. Flammini dataset, applying DenseNet169 and adaptability to other
preprocessing ResNet50 boosting datasets or real-world
techniques and accuracy. Tailored scenarios. The
transfer learning with preprocessing complexity of data
DenseNet169 and improves data quality. preprocessing and
ResNet50 for model transfer learning may
training. also increase
computational
demands.

11
 OBJECTIVES
This project aims to address the challenges in cervical cancer detection by:

1. To develop a deep learning model for extracting features from pap-smear images using light weight
models like Mobilenet, EfficientNet, and transformers like ViT improving the detection of normal, pre-
cancerous, and cancerous cells.
2. To enhance the model’s performance by applying Genetic Algorithm to optimize feature selection,
reducing dimensionality while improving classification accuracy.
3. To design and implement a new classification model to accurately categorize Pap-smear images,
leveraging optimized features and deep learning techniques to improve cervical cancer detection and
diagnosis.
4. To develop an interactive web application

12
 Dataset
The model is to be trained and tested upon the dataset:
SIPAKMED database

● The SIPaKMeD Database contains 4,049 images of isolated cervical cells, categorized into five
types: superficial-intermediate, parabasal, koilocytotic, dyskeratotic, and metaplastic cells. These
categories represent a range of normal, benign, and abnormal cell types seen in Pap smear tests. Each
cell type exhibits distinct morphological and staining characteristics, aiding in the identification of
cellular changes associated with cervical lesions. This database is a valuable resource for developing
automated diagnostic tools to assist in cervical cancer screening.

13
Dataset 14
Classes in sipakmed dataset

1. Superficial-Intermediate (SI)
● Description: These are mature squamous epithelial cells, usually found in normal cervical smear samples.

● Appearance: Large, flat cells with small, regular nuclei and ample cytoplasm.

● Clinical Significance: Indicates normal, healthy cells—commonly seen in reproductive-age women.

2. Parabasal (PB)
● Description: These are immature squamous cells, typically found when estrogen levels are low (e.g., postmenopause or after
childbirth).

● Appearance: Rounder cells with relatively large nuclei and less cytoplasm.

● Clinical Significance: While not necessarily abnormal, a large presence can indicate atrophic changes or hormonal
imbalances.

15
3. Koilocytotic (K)

● Description: Cells showing koilocytosis, which is a hallmark of HPV (Human Papillomavirus) infection.

● Appearance: Characterized by a perinuclear halo—a clear area around the nucleus.

● Clinical Significance: Considered pre-cancerous or abnormal; early detection is important for HPV-related changes.

4. Dyskeratotic (D)

● Description: Abnormal keratinized cells often associated with dysplasia or cancerous changes.

● Appearance: Irregular shapes, hyperchromatic (dark) nuclei, and dense cytoplasm.

● Clinical Significance: Suggests precancerous or cancerous conditions; requires immediate medical follow-up.

5. Metaplastic (M)

● Description: These are transformation zone cells, formed when columnar cells are replaced by squamous cells.

● Appearance: Variable in shape and size; show intermediate features between immature and mature cells.

● Clinical Significance: Usually benign but important to monitor as the transformation zone is where most cervical cancers begin.

16
Architecture

17
18
Modules Description

19
1. Deep Feature Extractors
The deep feature extractors are composed of three distinct pre-trained convolutional and transformer-based architectures:
Vision Transformer (ViT), EfficientNet, and MobileNet. Each model is responsible for extracting high-level discriminative
features from the input cytology image. ViT operates by dividing the image into fixed-size patches and processing them
through transformer blocks, allowing it to model long-range spatial dependencies. EfficientNet employs a compound scaling
method to balance depth, width, and resolution, producing compact yet powerful feature representations. MobileNet, on the
other hand, uses depthwise separable convolutions to reduce computational cost while maintaining accuracy, making it
suitable for efficient feature extraction. These three models together capture diverse visual characteristics, from fine-grained
texture details to abstract structural patterns.

2. Feature Concatenation Layer

Following the individual extraction process, the output feature vectors from ViT, EfficientNet, and MobileNet are
concatenated into a single unified feature representation. This feature fusion layer ensures that the combined vector
encapsulates complementary information from multiple receptive fields and architectures. The concatenated vector benefits
from the unique strengths of each backbone: global contextual awareness from ViT, balanced representation from
EfficientNet, and computational efficiency from MobileNet. This unified feature set serves as a robust input for subsequent
20
3. Genetic Algorithm for Feature Selection
Due to the high dimensionality of the fused feature vector, a Genetic Algorithm (GA) is utilized to perform optimal
feature selection. GA is a biologically inspired metaheuristic that evolves a population of candidate solutions
through selection, crossover, and mutation processes. In this context, each individual in the population represents a
binary mask determining which features to retain. The fitness function is typically based on the classification
performance (e.g., SVM 22
accuracy). By iteratively selecting the most informative and non-redundant features, GA improves generalization,
reduces overfitting, and enhances the classifier’s efficiency.

4. Support Vector Machine Classifier

The selected optimal features are finally passed to a Support Vector Machine (SVM), a powerful supervised
classifier that constructs a hyperplane in a high-dimensional space to separate the classes. SVM is particularly
effective in small-to-medium datasets with high-dimensional features, which is typical in medical image analysis.
The use of SVM ensures robust decision boundaries and high classification accuracy, making it suitable for
detecting subtle morphological differences in cervical cells. The final output corresponds to the predicted class
21
label of the cytology image (e.g., normal or abnormal cell types).
Result Analysis

22
23
The classification performance of the proposed hybrid model was evaluated using standard
metrics such as precision, recall, F1-score, and accuracy across all five cervical cell classes in
the SIPaKMeD dataset. The results are summarized in the classification report, which reflects
the robustness and reliability of the proposed approach.
● im_Dyskeratotic: Achieved a precision of 0.95 and a recall of 0.99, resulting in a strong
F1-score of 0.97, indicating that the model is highly effective at identifying dyskeratotic
cells with minimal false negatives.
● im_Koilocytotic: Obtained a precision of 0.97 and a recall of 0.92, with an F1-score of
0.94. While the precision is high, the slightly lower recall suggests a minor proportion of
missed positive instances for this class.

24
● im_Metaplastic: Recorded a precision and recall of 0.95 and 0.96, respectively,
yielding an F1-score of 0.96, confirming the model’s balanced performance in
classifying 25
● metaplastic cells.
● im_Parabasal: Scored 0.97 in both precision and recall, with a consistent F1-score of
0.97, demonstrating stable performance in recognizing this class.
● im_Superficial-Intermediate: Demonstrated exceptional performance with precision
and recall both at 0.99, and an F1-score of 0.99, indicating near-perfect classification
of these cell types.

25
 CONCLUSION
This project aims to address the challenges in cervical cancer detection by:

1. Developed a deep learning model for extracting features from pap-smear images using light weight
models EfficientNet and ViT improving the detection of normal, pre-cancerous, and cancerous cells.
2. Enhanced the model’s performance by applying Genetic Algorithm to optimize feature selection,
reducing dimensionality while improving classification accuracy.
3. Implemented a classification models like SVM to accurately categorize Pap-smear images, leveraging
optimized features and deep learning techniques to improve cervical cancer detection and diagnosis.

26
Contribution

27
Survey Paper

Status: Presented
First International Conference on Advances in Smart Computing and Applications, GLS
University, 2025
Date of Conference: 15th-16th February
Publisher: SpringerLink(SCOPUS Indexed)
28
Final Paper

Status: Communicated
2025 International Conference on Next Generation of Green Information and Emerging
Technologies, GIET University
Date of Conference: 8th and 9th August 2025
Publisher: IEEE Xplore(SCOPUS Indexed) 29
 REFERENCES
[1] S. Vachmanus, T. Noraset, W. Piyanonpong, T. Rattananukrom and S. Tuarob, "DeepMetaForge: A Deep Vision-
Transformer Metadata-Fusion Network for Automatic Skin Lesion Classification," in IEEE Access, vol. 11, pp. 145467-
145484, 2023, doi: 10.1109/ACCESS.2023.3345225.

[2] M. Fahaad Almufareh, "An Edge Computing-Based Factor-Aware Novel Framework for Early Detection and
Classification of Melanoma Disease Through a Customized VGG16 Architecture With Privacy Preservation and Real-Time
Analysis," in IEEE Access, vol. 12, pp. 113580-113596, 2024, doi: 10.1109/ACCESS.2024.3444050.

[3] S. Chatterjee, J. -M. Gil and Y. -C. Byun, "Early Detection of Multiclass Skin Lesions Using Transfer Learning-Based
IncepX-Ensemble Model," in IEEE Access, vol. 12, pp. 113677-113693, 2024, doi: 10.1109/ACCESS.2024.3432904.

[4] V. S. S. B. T. Sathvika et al., "Pipelined Structure in the Classification of Skin Lesions Based on Alexnet CNN and SVM
Model With Bi-Sectional Texture Features," in IEEE Access, vol. 12, pp. 57366-57380, 2024, doi:
10.1109/ACCESS.2024.3387533.

[5] X. Qian et al., "SPCB-Net: A Multi-Scale Skin Cancer Image Identification Network Using Self-Interactive Attention
Pyramid and Cross-Layer Bilinear-Trilinear Pooling," in IEEE Access, vol. 12, pp. 2272-2287, 2024, doi:
10.1109/ACCESS.2023.3347424.
30
[6] Y. Xie, J. Zhang, Y. Xia and C. Shen, "A Mutual Bootstrapping Model for Automated Skin Lesion Segmentation and
Classification," in IEEE Transactions on Medical Imaging, vol. 39, no. 7, pp. 2482-2493, July 2020, doi:
10.1109/TMI.2020.2972964.
keywords: {Lesions;Image segmentation;Skin;Task analysis;Feature extraction;Decoding;Computational modeling;Skin
lesion segmentation;skin lesion classification;deep convolutional neural network;dermoscopy},

[7] N. Nigar, M. Umar, M. K. Shahzad, S. Islam and D. Abalo, "A Deep Learning Approach Based on Explainable Artificial
Intelligence for Skin Lesion Classification," in IEEE Access, vol. 10, pp. 113715-113725, 2022, doi:
10.1109/ACCESS.2022.3217217.

[8] M. Saeed, A. Naseer, H. Masood, S. U. Rehman and V. Gruhn, "The Power of Generative AI to Augment for Enhanced
Skin Cancer Classification: A Deep Learning Approach," in IEEE Access, vol. 11, pp. 130330-130344, 2023, doi:
10.1109/ACCESS.2023.3332628.

[9] A. Magdy, H. Hussein, R. F. Abdel-Kader and K. A. E. Salam, "Performance Enhancement of Skin Cancer Classification
Using Computer Vision," in IEEE Access, vol. 11, pp. 72120-72133, 2023, doi: 10.1109/ACCESS.2023.3294974.

[10] H. L. Gururaj, N. Manju, A. Nagarjun, V. N. M. Aradhya and F. Flammini, "DeepSkin: A Deep Learning Approach for
Skin Cancer Classification," in IEEE Access, vol. 11, pp. 50205-50214, 2023, doi: 10.1109/ACCESS.2023.3274848.

31
Thank
You

32