Pneumonia Radiography Dataset
A set of medical images of pediatric chest x-rays, annotated by experts to detect viral or bacterial pneumonia.
Approximately 4,929 classified JPEG images (pneumonia vs normal), organized in train/test/val
CC0: Public Domain
Description
The dataset Pneumonia Radiography is a structured medical data set containing nearly 5,000 pediatric chest x-rays. The images are classified into two groups: normal cases and cases of pneumonia (bacterial or viral). The diagnoses were carried out by two radiologists, then validated by a senior expert, ensuring high reliability of the annotations.
What is this dataset for?
- Training classification models for the detection of pneumonia on medical images
- Evaluating the accuracy of convolutional neural networks (CNN) on public health tasks
- Testing medical image analysis pipelines in a pediatric setting
Can it be enriched or improved?
Yes, it is possible to increase this dataset by generating synthetic images, or by combining clinical metadata (age, symptoms). The addition of segments or masks could make it possible to train more accurate detection models (segmentation or location).
🔎 In summary
🧠 Recommended for
- Public health projects
- Machine learning applied to pediatrics
- Classification by medical vision
🔧 Compatible tools
- TensorFlow
- PyTorch
- Keras
- FastAI
- OpenCV
💡 Tip
Use this dataset to test different CNN architectures (ResNet, VGG) and compare their sensitivity on viral vs bacterial cases.
Frequently Asked Questions
Does this dataset contain segmentation or detection annotations?
No, the images are only labeled in binary classification (pneumonia or normal), without masks or bounding boxes.
Is this dataset suitable for direct clinical use?
No, it is intended for training experimental models, and should not be used in diagnosis without medical validation.
Can this dataset be used for multi-class detection (bacterial vs viral)?
Yes, although subcategories are mentioned, they are not separated into separate folders. They will have to be labelled manually.
