Pap Smear Cell Classification using Convolutional Neural Networks

This repository contains a Python-based machine learning project for automatic classification of Pap smear cell images into three clinically relevant classes using Convolutional Neural Networks (CNNs).

The project is intended for academic and educational purposes.

---

Project Overview

The goal of this project is to classify cervical cell images from Pap smear tests into three classes:

• Normal
• LG lesion (Low-grade lesion)
• HG lesion (High-grade lesion)

The model is trained using a CNN architecture implemented in TensorFlow/Keras and evaluated using standard classification metrics such as accuracy, precision, recall, F1-score, and confusion matrix.

---

Original Dataset Structure

The original dataset (Herlev Pap Smear Dataset) contains seven subclasses organized as follows:

smear2005/
└── New database pictures/
    ├── normal_superficiel/
    ├── normal_columnar/
    ├── normal_intermediate/
    ├── light_dysplastic/
    ├── moderate_dysplastic/
    ├── severe_dysplastic/
    └── carcinoma_in_situ/

These seven subclasses are mapped into three target classes:

• Normal

  • normal_superficiel
  • normal_columnar
  • normal_intermediate

• LG (Low-grade lesion)

  • light_dysplastic

• HG (High-grade lesion)

  • moderate_dysplastic
  • severe_dysplastic
  • carcinoma_in_situ

You can download dataset: Herlev Pap Smear Dataset https://mde-lab.aegean.gr/index.php/downloads/ (file: smear2005.zip)

• Reference Paper: https://telkomnika.uad.ac.id/index.php/TELKOMNIKA/article/view/22440

Dataset should be imported in project as data/raw/smear2005/New database pictures

---

Dataset Preparation

Dataset preprocessing and class mapping are performed using the script:

python src/prepare_dataset.py

This script:

• Reads images from the original dataset structure
• Maps the 7 subclasses into 3 classes (Normal, LG, HG)
• Copies images into a new processed dataset structure

Resulting directory structure:

data_processed/
├── Normal/
├── LG/
└── HG/

Dataset Loading and Verification

The dataset.py module:

• Loads images from data_processed/
• Resizes images to 128 × 128 pixels
• Normalizes pixel values to the range [0, 1]
• Prepares NumPy arrays suitable for CNN training

Quick sanity check

You can verify that dataset loading works correctly by running:

python

Then inside the Python shell:

from src.dataset import get_train_test_data

X_train, X_test, y_train, y_test = get_train_test_data()

print(X_train.shape)
print(y_train.shape)

Expected output:

(XXXX, 128, 128, 3)
(XXXX,)

Total dataset size:

(1834, 128, 128, 3)

---

CNN Model

The CNN architecture is defined in:

src/model.py

It includes:

• Convolutional layers for feature extraction
• Pooling layers for dimensionality reduction
• Fully connected layers
• Softmax output layer for 3-class classification

---

Model Training

Training logic is implemented in:

src/train.py

The script:

• Loads processed data using dataset.py
• Builds the CNN model using model.py
• Trains the model with early stopping regularization
• Saves the trained model to:

results/cnn_model.h5

---

Environment Setup

At the moment of posting repository TensorFlow does not support the latest Python versions. Therefore, Python 3.10 is used together with a virtual environment to avoid dependency conflicts. If your version of python supports it, you can skip this part.

Create and activate virtual environment

py -3.10 -m venv venv
venv\Scripts\activate

Install required dependencies

pip install tensorflow==2.15.0 opencv-python scikit-learn numpy

---

Training the Model

From the project root directory (with virtual environment activated):

python -m src.train

After training, the model will be saved as:

results/cnn_model.h5

---

Model Evaluation

To evaluate the trained model and visualize results:

Install additional libraries

pip install matplotlib seaborn

Run evaluation

python -m src.evaluate

This script outputs:

• Accuracy
• Precision, Recall, F1-score per class
• Confusion Matrix visualization

---

Notes

• The dataset is not included in this repository.
• This project is designed for educational and research purposes.
• Results may vary depending on hardware and random initialization.