To prepare the dataset in HDF5 format, run the following command:

python prepare_dataset.py

You can modify the dataset name in the configuration.txt file.

DRIVE, STARE, CHASE_DB, HRF

The workflow consists of two main steps:

1. Training an FCN (Fully Convolutional Network) model
2. Testing the trained FCN model

To start training, run the following command:

python pytorch_train.py

• Model architecture and training settings are configured in configuration.txt.
• Number of sub-images (N_subimgs) for different datasets:
  • DRIVE & STARE: 20,000
  • CHASE_DB1: 21,000
  • HRF: 30,000
  • Private dataset: 90,000
• Training parameters:
  • Epochs (N-epochs): 100
  • Batch size (batch_size): 35
  • Learning rate (lr): 3e-3

To test the trained model, run the following command:

python pytorch_predict_fcn.py

• Stride settings for testing:
  • DRIVE, STARE, CHASE_DB1: stride_height = 5, stride_width = 5
  • HRF, Private dataset: stride_height = 10, stride_width = 10

Our pretrained model used in paper are in GoogleDrive.

After completing all processes, run the following command to obtain evaluation results:

python evalution.py

I'm very grateful for my co-first author, Chee Hong Lee's(cheehong200292@gmail.com) diligent efforts and contributions. Many thanks for codes of these baseline backbone networks, including DUNet, DSCNet, AttUNet, UKAN, RollingUNet, MambaUNet, CTFNet, IterNet, BCDUNet, UNet++. Transforms refer to torchbiomed.