Pneumonia X-Ray Classifier (Binary Image Classification)

This project uses deep learning techniques to classify chest X-ray images into two categories: Pneumonia and Normal. It applies both a Custom Convolutional Neural Network (CNN) and a Transfer Learning approach (MobileNetV2) to compare performance on a real-world medical diagnosis task.

⚠️ Dataset not included in this repo due to large size — follow the instructions below to download and set it up locally.

---

📁 Project Structure

pneumonia-xray-classifier/
├── dataset/
│ ├── train/
│ │ └── >> Contains Images of NORMAL/ and PNEUMONIA/
│ └── test/
│ └── >> Contains Images of NORMAL/ and PNEUMONIA/
├── models/
│ └── >> Saved trained models (.keras)
├── nootbooks/
│ ├── 1_data_exploration.ipynb
│ ├── 2_train_custom_cnn.ipynb
│ ├── 3_train_transfer_model.ipynb
│ └── 4_evaluate_model.ipynb
├── .gitignore
├── README.md
└── requirements.txt

---

🧪 Dataset

• Name: Chest X-Ray Images (Pneumonia)
• Source: Kaggle Dataset
• Size: ~2 GB
• Classes: NORMAL, PNEUMONIA

🔽 Download Instructions

1. Visit the dataset page on Kaggle
2. Download and extract it
3. Place the dataset folder like this:
   dataset/
   └── train/
   └── test/

---

⚙️ Setup Instructions

1. Clone this repo:

   git clone https://github.com/your-username/pneumonia-xray-classifier.git
   cd pneumonia-xray-classifier

2. Create a virtual environment:

   python -m venv .venv
   .venv\Scripts\activate
   # or source .venv/bin/activate on Linux/MacOS

3. Install dependencies:

   pip install -r requirements.txt
   

4. Open notebooks:

   jupyter notebook
   

---

🧠 Models Used

✅ Custom CNN

• Built from scratch using Keras Sequential API
• Architecture includes:
  • 3 convolutional layers with ReLU activation
  • MaxPooling layers
  • Fully connected Dense layers
  • Dropout to reduce overfitting
• Output layer with sigmoid activation for binary classification

✅ MobileNetV2 (Transfer Learning)

• Pre-trained on ImageNet dataset
• Only the top layers (classification head) are trainable
• Fast convergence with better generalization
• Architecture:
  • GlobalAveragePooling
  • Dense layers with dropout
  • Sigmoid output layer

---

📈 Results (Sample)

Model	Accuracy	Precision	Recall	F1 Score
Custom CNN	0.76	0.72	0.99	0.836
MobileNetV2	0.83	0.79	0.99	0.880

🔬 These results are based on evaluation using the same test set. Actual values may vary depending on training epochs, hardware, and random initialization.

---

📄 License & Credits

• Dataset by Paul Mooney via Kaggle
• Transfer learning model: MobileNetV2 from TensorFlow/Keras
• Project developed for my AI&DL Assignment

---