Skip to main content

​
Dataset Overview

​
Chest X-Ray Images (Pneumonia)

Source: Kaggle
Author: Paul Mooney
URL: https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia
Size: ~1.15 GB (5,863 images)
License: CC BY 4.0

​
Description

The dataset contains chest X-ray radiographs from pediatric patients (ages 1-5 years) from the Guangzhou Women and Children’s Medical Center, China.
Images are organized into 3 folders (train, test, val) with subfolders for each category (Pneumonia/Normal).

​
Dataset Composition

​
Distribution Statistics

Total images: 5,863
├── Train:     5,216 images (89%)
│   ├── NORMAL:    1,341 images
│   └── PNEUMONIA: 3,875 images
├── Test:        624 images (11%)
│   ├── NORMAL:    234 images
│   └── PNEUMONIA: 390 images
└── Val:          16 images (very small, not used)
The original validation set contains only 16 images, which is too small for reliable validation. Our code automatically creates a proper 80/20 split from the training set.

​
Class Distribution

The dataset has class imbalance:
  • Normal cases: ~26% of training data
  • Pneumonia cases: ~74% of training data
This imbalance reflects real-world clinical scenarios but requires careful handling during training.

​
Download Instructions

1

Visit Kaggle

Go to https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumoniaIf you don’t have a Kaggle account, create one (it’s free)
2

Download the Dataset

Click the Download button (top right)A file named archive.zip (~1.15 GB) will download
3

Extract Files

Unzip archive.zipYou’ll find a folder named chest_xray
4

Move to Project

Place the extracted folder in your project:
proyectoia/
└── data/
    └── chest_xray/
        ├── train/
        │   ├── NORMAL/
        │   │   └── [1341 .jpeg images]
        │   └── PNEUMONIA/
        │       └── [3875 .jpeg images]
        └── test/
            ├── NORMAL/
            │   └── [234 .jpeg images]
            └── PNEUMONIA/
                └── [390 .jpeg images]

​
Option 2: Download with Kaggle API

For command-line enthusiasts:
1

Install Kaggle API

pipenv install kaggle
2

Configure Credentials

  1. Go to your Kaggle profile → Account → API → Create New Token
  2. Download kaggle.json
  3. On Windows, move it to: C:\Users\<username>\.kaggle\kaggle.json
  4. On Linux/Mac: ~/.kaggle/kaggle.json and run chmod 600 ~/.kaggle/kaggle.json
3

Download Dataset

pipenv run kaggle datasets download -d paultimothymooney/chest-xray-pneumonia -p data/
cd data
unzip chest-xray-pneumonia.zip -d chest_xray

​
Dataset Verification

After downloading, verify the structure: 
pipenv run python verificar_dataset.py

​
Expected Structure

data/chest_xray/
├── train/
│   ├── NORMAL/         [1,341 images]
│   └── PNEUMONIA/      [3,875 images]
├── test/
│   ├── NORMAL/         [234 images]
│   └── PNEUMONIA/      [390 images]
└── val/ (optional, not used)
If the verification script reports any missing images or incorrect counts, re-download the dataset to ensure data integrity.

​
Image Characteristics

​
Technical Specifications

  • Format: JPEG
  • Dimensions: Variable (resized to 224×224 in code)
  • Channels: 3 (RGB, though X-rays are grayscale)
  • Quality: Real clinical-use images
  • Bit Depth: 8-bit per channel

​
Clinical Categories

NORMAL
  • Clear lung parenchyma
  • No abnormal opacities
  • Normal cardiothoracic ratio
  • Clear costophrenic angles
PNEUMONIA
  • Presence of infiltrates
  • Opacities or consolidations
  • Increased density in lung fields
  • May show air bronchograms
The dataset includes both viral and bacterial pneumonia cases, though they are not separately labeled.

​
License Information

The dataset is released under CC BY 4.0 license, which permits:
  • ✅ Commercial and non-commercial use
  • ✅ Modification and distribution
  • ✅ Research and education
Requirement: Provide appropriate attribution to the creator.

​
Privacy Protection

  • All images have been anonymized
  • No patient-identifiable information is present
  • Approved for research use
  • Compliant with medical data sharing regulations

​
Responsible Use

CRITICAL: This model is for educational and research purposes ONLY.DO NOT use for:
  • Real clinical diagnosis without medical supervision
  • Replacing professional radiologist judgment
  • Medical decision-making without additional validation
  • Patient care without proper clinical oversight
AI systems are clinical decision support tools, not replacements for healthcare professionals.

​
Preprocessing Pipeline

Our code automatically applies the following preprocessing:

​
1. Resizing

  • All images resized to 224×224 pixels
  • Maintains aspect ratio considerations
  • Standard input size for the CNN

​
2. Normalization

  • Pixel values scaled to [0, 1] range
  • Divides by 255.0
  • Improves training stability

​
3. Data Augmentation (Training Only)

  • Random rotations: ±15 degrees
  • Simulates different patient positioning
  • Improves rotation invariance
  • Horizontal/vertical shifts: ±10%
  • Accounts for different X-ray centering
  • Enhances spatial robustness
  • Random zoom: ±10%
  • Simulates different patient distances
  • Improves scale invariance
  • Random horizontal flipping
  • Anatomically valid (lungs are roughly symmetric)
  • Doubles effective training data
Data augmentation is applied only during training, not during validation or testing, to ensure fair evaluation.

​
Dataset Split Strategy

​
Our Approach

Since the original validation set is too small, we:
  1. Combine train and val sets
  2. Create new split: 80% train / 20% validation
  3. Keep original test set separate and untouched

​
Final Distribution

Training:   ~4,173 images (80% of combined train+val)
Validation: ~1,043 images (20% of combined train+val)
Test:         624 images (original test set, never seen during training)
The test set remains completely independent to provide unbiased performance evaluation.

​
Citation and Attribution

​
Dataset Citation

Kermany, Daniel; Zhang, Kang; Goldbaum, Michael (2018),
"Labeled Optical Coherence Tomography (OCT) and Chest X-Ray Images for Classification",
Mendeley Data, v2.
http://dx.doi.org/10.17632/rscbjbr9sj.2

​
Scientific Publication

Kermany DS, Goldbaum M, et al. (2018).
"Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning".
Cell. 172(5):1122-1131.
https://doi.org/10.1016/j.cell.2018.02.010
When using this dataset in your work, please provide proper attribution to respect the creators’ contribution to open medical AI research.

​
Common Issues and Solutions

Solution: Try the manual download option instead of the API. Large files can be interrupted on unstable connections.
Solution: Re-download the dataset. The zip file may have been corrupted during download.
Solution: On Linux/Mac, run chmod 600 ~/.kaggle/kaggle.json to set proper permissions.
Solution: Reduce batch size in training configuration. The dataset is large, and 224×224 RGB images require significant memory.

​
Next Steps

With the dataset downloaded and understood, you’re ready to:
  1. Verify the data with the verification script
  2. Explore the images to understand the visual patterns
  3. Train the model using the prepared data pipeline
  4. Evaluate results on the test set
The preprocessing pipeline in the code will handle all the technical details of loading and augmenting the data.

Build docs developers (and LLMs) love

Get started for freeTalk to us