Evaluating the feasibility of using Generative Models to generate Chest X-Ray Data

Using GANs and Stable Diffusion to generate Chest Xray data points and evaluating them using convolutional classifiers.

This repository is the official implementation of Evaluating the feasibility of using Generative Models to generate Chest X-Ray Data

How to run

• Open notebooks in the order mentioned below (preferably on colab, for ease of use due to formatting)
• Descriptions and instructions are provided in the notebooks

Training

To train the models in the paper, open this notebook:

Chest XRay Image Synthesis The PGGAN is pre-trained and can only be used to generate samples, meanwhile the stable diffusion model is fine-tuned in the notebook and saved to huggingface. (Make a copy if you wish to make changes.)

Data Generation and Evaluation

To generate the chest x-ray data and evaluate using a classifier, open this notebook:

Data Synthesis For Image Classification This notebook is used to genarate multiple images to be used as a dataset or in addition to a dataset. (make a copy if you wish to make changes)

Processed Dataset

As mentioned in the paper, we have processed the Chest X-ray 14 dataset to prepare it for stable diffusion fine-tuning. The images can be found here

Pre-trained Models

You can download pretrained model here:

• Model Checkpoints trained on Chest Xray14 train set. (fine-tuned version of stable diffusion v2.1-512px)

Use the data generation and evalutation notebook linked above to use any of these models yourself.

Results

• Quantitative results:

  Classifier trained on no finding chest x-rays vs Edema chest x-rays, first with only real images and then with both real and synthetic images (synthetic images in the train set only)

  

Dataset

All our models and evaluations have utilised the Chest Xray14 dataset: ChestXray14

Other credits

• PGGAN Implementation: Segal, Bradley, et al. “Evaluating the Clinical Realism of Synthetic Chest x-Rays Generated Using Progressively Growing Gans.” SN Computer Science, vol. 2, no. 4, 2021, https://doi.org/10.1007/s42979-021-00720-7.
• Dreambooth: https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb
• Classification Model: https://www.kaggle.com/code/heyytanay/xray-image-eda-classification-keras/notebook