Pneumonia Detection from Chest X-Ray Images

A deep-learning web app that classifies a chest X-ray as Normal or Pneumonia using a Convolutional Neural Network (CNN). Upload an X-ray and get a prediction with a confidence score.

⚠️ Educational project — not a medical device. Do not use for diagnosis or treatment. See DISCLAIMER.md.

About

Pneumonia inflames the air sacs in one or both lungs and is diagnosed clinically, often with the help of a chest X-ray. CNNs are a strong fit for image classification, and this project trains one to recognise pneumonia patterns in X-rays as a learning exercise in applied deep learning.

Dataset

Chest X-Ray Images (Pneumonia) — Kaggle (~1.24 GB, ~5,860 images across train / val / test, classes NORMAL and PNEUMONIA).

Model

A compact Keras Sequential CNN trained on 75×75 grayscale X-rays:

Conv2D(64) → MaxPooling2D → Flatten → Dense(512) → Dense(256) → Dense(1, sigmoid)

Binary cross-entropy loss, Adam optimizer. The full pipeline is in CNN_model.ipynb, which saves the trained weights to cnn_model.h5.

Tech stack

Component	Technology
Model	TensorFlow / Keras (CNN)
Backend	Flask
Frontend	HTML, CSS, JavaScript (jQuery)
Image processing	NumPy, Pillow

Project structure

├── app.py                 # Flask web application
├── CNN_model.ipynb        # Model training notebook
├── requirements.txt       # Python dependencies
├── DISCLAIMER.md          # Medical disclaimer
├── tests/
│   └── test_smoke.py      # Smoke tests (TensorFlow mocked)
├── templates/
│   └── index.html         # Web interface
└── static/
    ├── css/  ├── js/  └── images/

Getting started

1. Install dependencies

python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Get a trained model

Either train it yourself — download the dataset and run CNN_model.ipynb (set DATA_DIR to your local dataset path); it saves cnn_model.h5 — or drop an existing cnn_model.h5 in the project root.

3. Run the web app

python app.py                 # serves on http://localhost:5000

Open http://localhost:5000, upload a chest X-ray, and click Predict.

API

Route	Method	Description
/	GET	Web UI
/health	GET	Liveness probe → {"status":"ok","model_present":bool}
/predict	POST	multipart image file → {"label","pneumonia","confidence"}

curl -F "image=@chest_xray.png" http://localhost:5000/predict
# {"label":"Pneumonia detected","pneumonia":true,"confidence":87.3}

Tests

Smoke tests mock TensorFlow, so they run in seconds without the weights or dataset:

pip install pytest && pytest -q

Limitations

• Trained only on the Kaggle dataset; performance on X-rays from other equipment, populations, or preprocessing pipelines is not guaranteed.
• 75×75 grayscale input discards fine detail a clinical model would retain.
• Binary output only (Normal vs Pneumonia) — it does not identify pneumonia type, severity, or location, and gives no uncertainty calibration.
• Not validated for clinical use. See DISCLAIMER.md.

License

MIT © Sanjay Santhanam