This repository is the official PyTorch implementation of paper: Dual-Branch Network with Dual-Sampling Modulated Dice Loss for Hard Exudate Segmentation in Colour Fundus Images .

• The dual-HED folder contains vanilla and DualNet versions of HED.
• The dual-segmentation-toolbox folder contains vanilla and DualNet versions of PSPNet and Deeplabv3.
• The preprocessing and evaluation scripts can be found in scripts folder.

• Python: 3.7

• PyTorch: 1.1.0

• The dual-segmentation-toolbox code also needs apex and inplace-abn.

conda create -n dualnet python=3.7 -y
conda activate dualnet

pip install torch==1.1.0 torchvision==0.3.0

cd apex
python setup.py install --cpp_ext
pip install inplace_abn==1.0.12

pip install opencv-python
pip install tqdm scipy scikit-image

# for evaluation scripts
pip install pandas sklearn
pip install cupy-cuda100

The ImageNet pretrained weights of backbones are available at this link.

Please refer to train.sh and eval.sh.

We evaluate our DualNet on DDR and IDRiD datasets.

The corresponding trained weights of our DualNet models are available at this link.

In the paper, we reported average performance over three repetitions, while our code only reported the best one among them.

This code is heavily borrowed from HED , pytorch-segmentation-toolbox, and BBN. Thanks for their contributions.

If you find this code useful in your research, please consider citing:

@article{liu2022dual,
  title={Dual-Branch Network With Dual-Sampling Modulated Dice Loss for Hard Exudate Segmentation in Color Fundus Images},
  author={Liu, Qing and Liu, Haotian and Zhao, Yang and Liang, Yixiong},
  journal={IEEE Journal of Biomedical and Health Informatics},
  volume={26},
  number={3},
  pages={1091--1102},
  year={2022},
  publisher={IEEE}，
  doi={10.1109/JBHI.2021.3108169}
}