Diffusion-based generative model for synthesising chest X-ray images using a 3D U-Net-style architecture and sinusoidal noise embeddings.
This repository implements a denoising diffusion probabilistic model (DDPM) adapted for medical imaging, specifically chest X-ray generation.
The project demonstrates how diffusion models originally developed for natural images can be extended to a healthcare domain.
The pipeline includes:
- dataset loading and preprocessing
- diffusion noise scheduling
- sinusoidal timestep embeddings
- U-Net denoising network
- exponential moving average (EMA) stabilisation
- image generation and visualisation
Output following 1 epoch:
Output following 25 epochs:
Output following 50 epochs:
Output following 100 epochs:
src/
config.py
data.py
diffusion.py
model.py
train.py
generate.py
utils.py
notebooks/
demo_generation.ipynb
Install dependencies:
- pip install -r requirements.txt
Train the model:
- python src/train.py
Generate X-ray images from a trained model:
- python src/generate.py
Diffusion-based generative modelling (DDPM-style) Sinusoidal timestep embeddings U-Net denoising architecture EMA stabilisation of weights Config-driven training pipeline Reproducible training and generation scripts
The dataset is not included in this repository, but is publicly available on Kaggle under an MIT license:
👉 https://www.kaggle.com/datasets/divyam6969/chest-xray-pneumonia-dataset
Expected structure:
data/
└── chest_xray/
└── train/
└──bacterial_pneumonia
└──viral_pneumonia
└──fungal_pneumonia
After downloading, place the dataset in the data/ directory as shown above.
Parts of this implementation are adapted from diffusion model examples presented in Foster Generative AI by David Foster, originally applied to natural image datasets (e.g., flowers).
This repository extends and modifies those approaches for medical imaging applications, including changes to dataset handling, model configuration, and training workflow.
Foster, D. Generative Deep Learning. 2nd Ed. O’Reilly Media.
Demonstrates application of diffusion models to medical image generation Translates generative AI methods from natural images to healthcare Provides a clean, modular implementation for experimentation and extension
David Power
Healthcare Simulation Specialist | MSc Artificial Intelligence
- 💼 LinkedIn: https://www.linkedin.com/in/dave-power-47280a44/
- 💻 GitHub: https://github.com/DavePower-cloud