Pneumonia Detection from Chest X-ray Images using Transfer Learning

This project demonstrates a deep learning model for classifying chest X-ray images as either 'NORMAL' or 'PNEUMONIA'. It leverages transfer learning with the VGG16 architecture to achieve high accuracy and recall.

The code is structured for use in a Google Colab environment and implements a strategic two-phase training process to maximize the performance of the pre-trained model on this specific medical imaging task. This project was completed as part of my internship to showcase skills in computer vision, deep learning model development, and experimental iteration.

---

Key Features

• Utilizes Transfer Learning with the powerful, pre-trained VGG16 model from ImageNet.
• Implements a two-phase training strategy for optimal results:
  1. Initial Training (Feature Extraction): Training a custom classifier head on top of a frozen base model to learn high-level features.
  2. Fine-Tuning: Unfreezing the VGG16 base model and training the entire network with a low learning rate to adapt it to the specifics of X-ray images.
• Handles the dataset's class imbalance using calculated class weights to ensure the model learns fairly from both classes.
• Employs data augmentation (random rotations, shifts, zooms, flips) to create a more robust model that generalizes better to new, unseen images.
• Achieved a final validation accuracy of ~93.4% with high precision and recall.

---

Dataset

The model was trained on the "Chest X-ray Images (Pneumonia)" dataset available on Kaggle.

• Link to Dataset: https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia

---

Methodology

The core of this project is the two-phase training approach, which is a standard and effective technique in transfer learning:

1. Initial Phase (Training the Head): We start with the VGG16 model, with its weights pre-trained on ImageNet, and freeze the entire base. We then add our own small classifier head (consisting of a GlobalAveragePooling2D, Dropout, and a final Dense layer with a sigmoid activation function). This new head is trained for several epochs on our X-ray data. This allows our new layers to learn how to interpret the features extracted by the powerful, frozen VGG16 base without disrupting its valuable, pre-learned weights. A very low learning rate is used to ensure stable learning.

2. Fine-Tuning Phase: After the initial phase provides a good set of weights for the classifier head, we unfreeze the VGG16 base model. The entire network is then trained with a very low learning rate (5e-5). This allows the model to make small, careful adjustments to the pre-trained features to better suit the specific nuances of the chest X-ray images, leading to a significant performance boost.

Callbacks like EarlyStopping and ReduceLROnPlateau are used in both phases to prevent overfitting, save the best-performing model, and optimize the learning rate.

---

Final Performance

The final model, after the completion of the fine-tuning phase, achieved the following performance on the validation set (624 images). The weights from the best-performing epoch were restored by the EarlyStopping callback.

• Validation Accuracy: ~93.43%
• Validation Loss: ~0.1887
• Precision (for PNEUMONIA): ~94.86% (When the model predicts PNEUMONIA, it is correct ~95% of the time).
• Recall (for PNEUMONIA): ~94.62% (The model correctly identifies ~95% of all actual PNEUMONIA cases).
• F1-Score (for PNEUMONIA): ~94.74%

---

How to Use This Project

1. Setup

• Clone the repository:

  git clone [https://github.com/](https://github.com/)[Your-GitHub-Username]/[Your-Repo-Name].git
  cd [Your-Repo-Name]

• Install dependencies:

  pip install -r requirements.txt

• Download the dataset:
  • The training notebook (Pneumonia_Detection_Training.ipynb) contains a cell to download the dataset from Kaggle. You will need a kaggle.json API token file from your Kaggle account.
  • Place the downloaded and unzipped data in a directory structure that the script expects, e.g., /content/datasets/pneumonia/chest_xray/ with train and test subfolders.

2. Training the Model

• The primary training code is in the Jupyter Notebook (Pneumonia_Detection_Training.ipynb).
• Open the notebook in an environment like Google Colab or Jupyter Lab.
• Run the cells sequentially to download the data, build the model, and execute the two-phase training process.
• The final trained model will be saved as pneumonia_detection_finetuned_fixed.keras.

3. Making Predictions on New Images

• A script predict_pneumonia.py is provided to load the saved model and make predictions on new images.
• Place your sample X-ray images in the sample_images/ directory.
• Run the script from your terminal:

  python predict_pneumonia.py

• The script will load the .keras model, process each image, and print the predicted class ('NORMAL' or 'PNEUMONIA') along with the model's confidence.

---

Project Structure

. ├── Pneumonia_Detection_Training.ipynb # Main notebook for training the model ├── predict_pneumonia.py # Script to predict on new images ├── requirements.txt # List of Python dependencies ├── pneumonia_detection_finetuned_fixed.keras # The final saved model file ├── sample_images/ # Directory for placing sample images for prediction │ ├── normal_sample.jpeg │ └── pneumonia_sample.jpeg └── README.md # This file

---

Technologies Used

• Python 3
• TensorFlow & Keras
• Scikit-learn
• NumPy
• Matplotlib
• Kaggle API