Skip to content

Chest X-Ray Datasets

This document covers chest X-ray datasets available in genai-lab for training diffusion models and other generative architectures.

1. Synthetic X-Ray Dataset

Purpose: Testing and development without requiring real medical data.

Description

SyntheticXRayDataset generates diverse synthetic chest X-ray images with randomized anatomical structures:

  • Lungs: Elliptical shapes with variable position, size, darkness
  • Heart: Circular structure between lungs
  • Ribs: Curved horizontal lines with variable spacing
  • Spine: Optional vertical structure
  • Clavicles: Optional horizontal structures at top

Usage

from genailab.diffusion.datasets import SyntheticXRayDataset

# Create dataset
dataset = SyntheticXRayDataset(
    n_samples=1000,    # Number of images to generate
    img_size=128,      # Image dimensions (square)
    seed=42            # Reproducibility
)

# Access images
img = dataset[0]  # Returns tensor of shape (1, 128, 128) in [-1, 1]

Characteristics

Property Value
Output shape (1, H, W) grayscale
Value range [-1, 1] (normalized)
Generation On-init (all images pre-generated)
Randomization Per-image anatomical variation

When to Use

  • Development: Fast iteration without data download
  • Testing: Verify pipeline correctness
  • Debugging: Reproducible synthetic data
  • Demos: Quick demonstrations

2. Real Chest X-Ray Dataset (Kaggle)

Purpose: Training on real medical images for realistic diffusion models.

Source

Kaggle Chest X-Ray Images (Pneumonia)

  • URL: https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia
  • License: CC BY 4.0
  • Size: ~1.2 GB

Dataset Structure

chest_xray/
├── train/
│   ├── NORMAL/      (~1,300 images)
│   └── PNEUMONIA/   (~3,900 images)
├── val/
│   ├── NORMAL/
│   └── PNEUMONIA/
└── test/
    ├── NORMAL/
    └── PNEUMONIA/

Usage

from genailab.diffusion.datasets import ChestXRayDataset

# Load real X-rays (normal cases)
dataset = ChestXRayDataset(
    root_dir="data/chest_xray/train/NORMAL",
    img_size=128
)

# Or use factory function
from genailab.diffusion.datasets import get_dataset

dataset = get_dataset(
    name='chest_xray',
    root_dir="data/chest_xray/train/NORMAL",
    img_size=128
)

Download Instructions

  1. Create Kaggle account and API token
  2. Install kaggle CLI: pip install kaggle
  3. Download:
kaggle datasets download -d paultimothymooney/chest-xray-pneumonia
unzip chest-xray-pneumonia.zip -d data/

Characteristics

Property Value
Original resolution Variable (typically 1000-2000 px)
Output shape (1, H, W) grayscale (resized)
Value range [-1, 1] (normalized)
Classes NORMAL, PNEUMONIA

3. Factory Function

Use get_dataset() for unified access:

from genailab.diffusion.datasets import get_dataset

# Synthetic
synthetic_ds = get_dataset('synthetic', img_size=128, n_samples=1000)

# Real
real_ds = get_dataset('chest_xray', root_dir='data/chest_xray/train/NORMAL', img_size=128)

4. Integration with Training

With Diffusion Training

from torch.utils.data import DataLoader
from genailab.diffusion.datasets import get_dataset
from genailab.diffusion.training import train_image_diffusion

# Create dataset and loader
dataset = get_dataset('synthetic', img_size=64, n_samples=500)
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

# Train diffusion model
model, losses = train_image_diffusion(
    dataloader=dataloader,
    img_size=64,
    n_epochs=50,
    device='cuda'
)

See Also

  • notebooks/diffusion/03_medical_imaging_diffusion/ — Full tutorial

5. Extending with New Datasets

To add a new medical imaging dataset:

  1. Create a new Dataset class in src/genailab/diffusion/datasets.py
  2. Implement __len__ and __getitem__
  3. Ensure output is (C, H, W) tensor in [-1, 1]
  4. Add to get_dataset() factory function
  5. Document here
File Contents
src/genailab/diffusion/datasets.py SyntheticXRayDataset, ChestXRayDataset, get_dataset
src/genailab/diffusion/training.py train_image_diffusion
src/genailab/diffusion/architectures.py UNet2D for image diffusion