Deep Learning-Based Cervical Cancer
Classification Using Transfer Learning
Author: Ishaan Bhattacharya
Affiliation: RCC Institute of information and Technology
Abstract
Cervical cancer remains a leading health concern worldwide, requiring timely and accurate
diagnosis. In this study, we propose a robust image classification pipeline utilizing transfer
learning with ResNet50 to classify cervix images into five categories. Leveraging a balanced
dataset of 25,000 labeled images across cervix_dyk, cervix_koc, cervix_nya, cervix_pcr,
and cervix_swa, our model achieved an impressive 99% validation accuracy. The proposed
method demonstrates high precision, recall, and F1 scores, indicating near-perfect classification
performance suitable for healthcare applications.
Keywords
Deep Learning, Transfer Learning, ResNet50, Cervical Cancer, Image Classification, Medical
Imaging
Introduction
Cervical cancer screening through visual inspection of cervix images is challenging due to
variability in image quality, lighting conditions, and anatomical differences. Deep learning
approaches have shown promise in automating this process with high precision. This project
aims to improve diagnosis efficiency and reliability by implementing a pre-trained ResNet50
model fine-tuned on a large, balanced dataset of cervical images.
Motivation and Contribution
The motivation stems from the need for scalable and accurate diagnostic tools in resource-
constrained settings. Our contribution includes:
Design of a streamlined classification pipeline based on ResNet50
Comprehensive evaluation on a well-structured multiclass dataset
Performance optimization to reduce training time
Suggestions for real-world deployment strategies
Material and Methods
3.1 Data Pre-processing and Augmentation
Image resizing to match ResNet50 input dimensions
Normalization and augmentation techniques (rotation, flipping) applied
Dataset split into 80% training and 20% validation
3.2 Proposed Method
Base model: ResNet50 with pre-trained ImageNet weights
Custom classifier head with softmax activation for 5 output classes
Categorical cross-entropy loss, Adam optimizer, and early stopping
Model checkpointing and learning rate scheduling for stability
Experimental Results and Discussions
4.1 Experimental Setup
Platform: Google colabs with runtime Environment set to GPU t4
Framework: TensorFlow/Keras
Training configuration: ~10 seconds per epoch due to optimizations
4.2 Dataset Preparation
25,000 labeled images in five class folders
Equal representation across classes confirmed
4.3 Evaluation Metrics
Accuracy, Precision, Recall, F1-Score
Class-wise analysis confirms uniformly high scores
4.4 Performance Analysis
Class Precision Recall F1-Score
cervix_dyk 0.99 0.99 0.99
cervix_koc 0.99 0.99 0.99
cervix_nya 0.99 0.99 0.99
cervix_pcr 0.99 0.99 0.99
cervix_swa 0.99 0.99 0.99
✅ Found dataset path: /kaggle/input/multi-cancer/Multi Cancer/Multi
Cancer/Cervical Cancer
Class folders: ['cervix_koc', 'cervix_dyk', 'cervix_pab', 'cervix_sfi',
'cervix_mep']
Found 25000 files belonging to 5 classes.
Using 20000 files for training.
Found 25000 files belonging to 5 classes.
Using 5000 files for validation.
Detected classes: ['cervix_dyk', 'cervix_koc', 'cervix_mep', 'cervix_pab',
'cervix_sfi']
Epoch 1/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 78s 93ms/step - accuracy: 0.8361 - loss:
0.5223 - val_accuracy: 0.9586 - val_loss: 0.1285
Epoch 2/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 12ms/step - accuracy: 0.9694 - loss:
0.0971 - val_accuracy: 0.9736 - val_loss: 0.0843
Epoch 3/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 11ms/step - accuracy: 0.9813 - loss:
0.0540 - val_accuracy: 0.9664 - val_loss: 0.1479
Epoch 4/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 11s 11ms/step - accuracy: 0.9860 - loss:
0.0456 - val_accuracy: 0.9808 - val_loss: 0.0656
Epoch 5/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 11s 12ms/step - accuracy: 0.9862 - loss:
0.0450 - val_accuracy: 0.9524 - val_loss: 0.3571
Epoch 6/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 6s 10ms/step - accuracy: 0.9856 - loss:
0.0570 - val_accuracy: 0.9860 - val_loss: 0.0480
Epoch 7/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 12ms/step - accuracy: 0.9958 - loss:
0.0147 - val_accuracy: 0.9826 - val_loss: 0.0634
Epoch 8/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 10ms/step - accuracy: 0.9931 - loss:
0.0272 - val_accuracy: 0.9738 - val_loss: 0.1456
Epoch 9/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 12ms/step - accuracy: 0.9886 - loss:
0.0377 - val_accuracy: 0.9842 - val_loss: 0.0775
Epoch 10/10
625/625 ━━━━━━━━━━━━━━━━━━━━ 7s 11ms/step - accuracy: 0.9939 - loss:
0.0217 - val_accuracy: 0.9904 - val_loss: 0.0487
✅ Model saved to: /content/drive/MyDrive/resnet50_multicancer.keras
157/157 ━━━━━━━━━━━━━━━━━━━━ 14s 45ms/step
4.5 Discussion
Despite near-perfect metrics, further steps such as Grad-CAM explainability, confusion matrix
visualization, and cross-validation are suggested to improve clinical interpretability and trust.
Future work may include experimenting with EfficientNet and Vision Transformers, ensembling
models, and deploying mobile-friendly interfaces.
Conclusion
This study presents a high-accuracy deep learning model for cervical cancer classification using
ResNet50. With strong metrics and efficient training, it offers practical potential for healthcare
deployment. Further enhancements in interpretability and model robustness could make it
suitable for clinical settings.
Acknowledgement
This work has been partially supported by the MeitY Project, MHRD, Govt. of India.
I am deeply thankful to my teachers, Dr. Neelotpal Chakraborty and Dr. Lidia Ghosh for their
constant encouragement and insightful advice. I also acknowledge Kaggle for providing the
Cervical Cancer dataset and Google Colab for offering the computational resources that enabled
me to successfully carry out model training and experimentation.
Conflict of Interest
The authors hereby state that there is no conflict of interest.
Declaration of Interest
The authors declare that they have no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this article.
References
Libraries & Frameworks
TensorFlow & Keras (Modeling & Preprocessing)
[Link]
[Link]
KaggleHub (for accessing Kaggle datasets)
[Link]
Matplotlib & Seaborn (Visualization)
[Link]
[Link]
Scikit-learn (Confusion matrix & classification report)
[Link]
[Link].classification_report.html
Model Architecture
ResNet50 - Deep Residual Learning for Image Recognition
[Link]
Implementation: [Link]
Dataset
Multi-Cancer Medical Image Dataset
Source: Kaggle
URL: [Link]