---
title: RetinaGAN
emoji: 😻
colorFrom: gray
colorTo: red
sdk: streamlit
sdk_version: 1.20.0
app_file: app.py
pinned: false
license: mit
---

# RetinaGAN

Code Repository for: [**High-Fidelity Diabetic Retina Fundus Image Synthesis from Freestyle Lesion Maps**](https://opg.optica.org/abstract.cfm?uri=boe-14-2-533)

## About

RetinaGAN a two-step process for generating photo-realistic retinal Fundus images based on artificially generated or free-hand drawn semantic lesion maps.

![](assets/RetinaGAN_pipeline.png) 

StyleGAN is modified to be conditional in to synthesize pathological lesion maps based on a specified DR grade (i.e., grades 0 to 4). The DR Grades are defined by the International Clinical Diabetic Retinopathy (ICDR) disease severity scale; no apparent retinopathy, {mild, moderate, severe} Non-Proliferative Diabetic Retinopathy (NPDR), and Proliferative Diabetic Retinopathy (PDR). The output of the network is a binary image with seven channels instead of class colors to avoid ambiguity.

![](assets/cStyleGAN.png) 

The generated label maps are then passed through SPADE, an image-to-image translation network, to turn them into photo-realistic retina fundus images. The input to the network are one-hot encoded labels.

![](assets/GauGAN.png) 

## Usage

Download model checkpoints (see [here](checkpoints/README.md) for details) and run the model via Streamlit. Start the app via `streamlit run web_demo.py`.

## Example Images

Example retina Fundus images synthesised from Conditional StyleGAN generated lesion maps. Top row: synthetically generated lesion maps based on DR grade by Conditional StyleGAN. Other rows: synthetic Fundus images generated by SPADE. Images are generated sequentially with random seed and are **not** cherry picked.

| grade 0                                                      | grade 1                                                      | grade 2                                                      | grade 3                                                      | grade 4                                                      |
|--------------------------------------------------------------|--------------------------------------------------------------|--------------------------------------------------------------|--------------------------------------------------------------|--------------------------------------------------------------|
| ![](assets/sample_images/mask_class_0_batch_0.png)           | ![](assets/sample_images/mask_class_1_batch_0.png)           | ![](assets/sample_images/mask_class_2_batch_0.png)           | ![](assets/sample_images/mask_class_3_batch_0.png)           | ![](assets/sample_images/mask_class_4_batch_0.png)           |
| ![](assets/sample_images/image_class_0_batch_0_sample_0.png) | ![](assets/sample_images/image_class_1_batch_0_sample_0.png) | ![](assets/sample_images/image_class_2_batch_0_sample_0.png) | ![](assets/sample_images/image_class_3_batch_0_sample_0.png) | ![](assets/sample_images/image_class_4_batch_0_sample_0.png) |
| ![](assets/sample_images/image_class_0_batch_0_sample_1.png) | ![](assets/sample_images/image_class_1_batch_0_sample_1.png) | ![](assets/sample_images/image_class_2_batch_0_sample_1.png) | ![](assets/sample_images/image_class_3_batch_0_sample_1.png) | ![](assets/sample_images/image_class_4_batch_0_sample_1.png) |

## Cite this work

If you find this work useful for your research, give us a kudos by citing:

```
@article{hou2023high,
  title={High-fidelity diabetic retina fundus image synthesis from freestyle lesion maps},
  author={Hou, Benjamin},
  journal={Biomedical Optics Express},
  volume={14},
  number={2},
  pages={533--549},
  year={2023},
  publisher={Optica Publishing Group}
}
```