AI-Driven Pneumonia Detection Using CNNs and VGG16

Project Overview

This project leverages Convolutional Neural Networks (CNNs) to classify chest X-ray images as either Normal or Pneumonia. It encompasses data pre-processing, model development (including a custom CNN and transfer learning with VGG16), training, evaluation, and performance analysis.

---

Motivation

Pneumonia is a significant health concern worldwide, especially in children and the elderly. Early and accurate detection through chest X-rays is vital for effective treatment. This project aims to:

• Automate pneumonia detection using deep learning techniques.
• Compare the efficacy of a custom-designed CNN with a pre-trained VGG16 model.
• Provide insights into the application of CNNs in medical image classification.

---

Dataset Description

• Source: The dataset is sourced from the Chest X-Ray Images (Pneumonia) dataset on Kaggle.
• Structure:
  • Training Set: 5,216 images (3,875 Pneumonia, 1,341 Normal)
  • Validation Set: 16 images (all Normal)
  • Test Set: 624 images (390 Pneumonia, 234 Normal)
• Pre-processing:
  • Images were resized to 224x224 pixels.
  • Applied normalisation to scale pixel values.
  • Data augmentation techniques (e.g., rotation, flipping) were employed to enhance model generalisation.

---

Model Architectures

1. Custom CNN

• Architecture:
  • Three convolutional layers with ReLU activation and max-pooling.
  • Two fully connected layers leading to a softmax output.
• Regularisation:
  • Dropout layers to prevent overfitting.
• Optimizer:
  • Adam optimizer with a learning rate of 0.001.

2. VGG16 Transfer Learning

• Modification:
  • Utilized the VGG16 model pre-trained on ImageNet.
  • Replaced the top classifier layers with a custom fully connected network suitable for binary classification.
• Training Strategy:
  • Fine-tuned the top layers while keeping the convolutional base frozen initially, then unfroze some layers for further training.

---

Training Procedure

• Hyperparameters:
  • Batch size: 32
  • Epochs: 25
• Loss Function:
  • Binary Cross-Entropy Loss
• Data Augmentation:
  • Applied random rotations, shifts, and flips to augment the training data.
• Validation:
  • Monitored validation loss to implement early stopping and prevent overfitting.

---

Evaluation Results

• Custom CNN:
  • Accuracy: 85%
  • Recall (Sensitivity): 88%
  • Precision: 83%
  • F1-Score: 85%
• VGG16 Transfer Learning:
  • Accuracy: 90%
  • Recall (Sensitivity): 92%
  • Precision: 89%
  • F1-Score: 90%

Note: Replace the above metrics with actual results from your experiments.

---

Confusion Matrix

	Predicted Normal	Predicted Pneumonia
Actual Normal	200	34
Actual Pneumonia	31	359

Note: Replace the above values with actual results from your experiments.

---

Installation

1. Clone the Repository

git clone https://github.com/drnsmith/pneumonia-detection-CNN.git
cd pneumonia-detection-CNN

2. Install Dependencies

To install the required dependencies, run the following command:

pip install -r requirements.txt

---

Usage

1. Prepare the Dataset

Download the dataset from Kaggle. Place it in the data/ directory, organised into the following subdirectories:

data/
├── train/
├── val/
└── test/

2. Run Pre-processing

Execute the preprocess.py script to resize images and apply data augmentation:

python preprocess.py

3. Train the Model

Use the train.py script to train either the custom CNN or the VGG16 model:

python train.py

4. Evaluate the Model

Run the evaluate.py script to assess model performance on the test set:

python evaluate.py

---

Future Enhancements

1. Hyperparameter Tuning

Experiment with different learning rates, batch sizes, and optimisers to improve performance.

2. Additional Data

Incorporate more diverse datasets to enhance model robustness and generalisation.

3. Model Interpretability

Implement techniques like Grad-CAM or similar methods to visualise regions of interest in X-ray images and improve model explainability.

---

Productisation

This pneumonia detection system can be transformed into a cloud-based AI diagnostic tool for healthcare providers, offering real-time analysis of chest X-rays to support radiologists in early pneumonia detection. By integrating with hospital information systems (HIS) or telemedicine platforms, the model can provide automated pre-screening, reducing workload and improving diagnostic speed. Additional features like explainability with Grad-CAM visualisations can enhance trust and usability in clinical settings.

Monetisation

The system can be monetised through a subscription-based SaaS model, allowing clinics and hospitals to access AI-powered X-ray diagnostics on demand. Licensing the model as an API for integration into existing radiology software offers another revenue stream. Partnerships with telehealth companies and medical research institutions can further expand market reach, with premium offerings like custom model tuning for specific datasets or regulatory compliance adaptations.

---

Contributing

We welcome contributions to enhance the functionality and performance of this system. Please fork the repository and submit a pull request with your proposed changes.

---

Repository History Cleaned

As part of preparing this repository for collaboration, its commit history has been cleaned. This action ensures a more streamlined project for contributors and removes outdated or redundant information in the history.

The current state reflects the latest progress as of 24/01/2025.

For questions regarding prior work or additional details, please contact the author.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.