UWAFA-GAN

This code is the pytorch implementation of the UWAFA-GAN which has been accepted by JBHI 2024 named "UWAFA-GAN: Ultra-Wide-Angle Fluorescein Angiography Transformation via Multi-scale Generation and Registration Enhancement". It can be used to turning UWF scanning laser ophthalmoscopy(UWF-SLO) to the UWF fluorescein angiography(UWF-FA) and display the tiny vascular lesion areas and could be trained on a little misalignment paired UWF-SLO and UWF-FA.

Generation UWF-FA from UWF-SLO with nice detail

Pre-requisties

• Linux
• python>=3.7
• NVIDIA GPU (memory>=10G) + CUDA cuDNN

How to train

Preparation of the data

We are diligently navigating through ethical and approval procedures to provide large quantities of trainable data that can enhance the comprehensiveness of our open-source code. However, at present, we are unable to publicly release substantial volumes of data. We appreciate your understanding regarding this matter. However, if you have your own data, you can still train and generate using our code. We assume your data, with suffix as ".jpg", are in the path "dataset/yours", firstly we need to random crop them into the training dataset.

python utils/advan_random_crop.py --datadir dataset/yours  --output_dir dataset/data_slo2ffa --suffix .jpg --index_interval 0 --index_interval 100 

This will crop the 100 pairs suffix-jpg images from path "dataset/yours" and put the result into the path "dataset/data_slo2ffa". Then the training procedure can be exerted. Find the yaml file in the path "config/train_config.yaml", making sure the data_path of it is correct.

Part of train_config.yaml

batchsize: 4
epoch: 40
num_D: 2
n_layers: 4

# validation setting
validation_epoch: 41
val_dir: ''

# dataloader
data_path: ["dataset/data_slo2ffa"]
val_length: 900
seed: -1 
img_size: [832, 1088]

and run the command:

python -u train_changed.py

Get start to the evaluation

We provide 6 pairs of examples for evaluation in dataset/example_pairs, if you have only UWF-SLO and you can try it as well. To evaluate, the first UWF-SLO should be named as 1.jpg, the second as 2.jpg... Just like it in dataset/example_pairs

checkpoint configure

Download the exp_final and make sure the path "./weights/exp_final" correct. Directory "exp_final" should contain three elements:

├── exp_final
    ├──discriminator.pt
    ├──generator.pt
    ├──reg.pt

and run the command:

python -u utils/Model_evaluation_without_ffa.py --updir dataset/example_pairs \
    --model_updir weights/exp_final

The result will be saved in the path dataset/example_pairs

Citation

@article{ge2024uwafa,
  title={UWAFA-GAN: Ultra-Wide-Angle Fluorescein Angiography Transformation via Multi-scale Generation and Registration Enhancement},
  author={Ge, Ruiquan and Fang, Zhaojie and Wei, Pengxue and Chen, Zhanghao and Jiang, Hongyang and Elazab, Ahmed and Li, Wangting and Wan, Xiang and Zhang, Shaochong and Wang, Changmiao},
  journal={IEEE Journal of Biomedical and Health Informatics},
  year={2024},
  publisher={IEEE}
}