datasets

Synthetic Dataset

We provide 500 synthetic training samples at ./images/, their corresponding labels at ./labels/, and the vessel graphs their based on at ./vessel_graphs. The code for vessel graph generation can be found under ../vessel_graph_generation. To generate [N] new samples as used in the paper use the docker script:

docker run -v [RESULT_DIR]:/var/generation octa-seg generation [N]

You can also generate your own synthetic samples with a customized config (e.g. ./vessel_graph_generation/configs/dataset_18_June_2023.yml). Run from project folder:

# This will generate 500 samples at `./vessel_graph_generation/datasets/dataset_18_June_2023/`
python generate_vessel_graph.py --config_file ./vessel_graph_generation/configs/dataset_18_June_2023.yml --num_samples 500

To visualize them run:

# Create images
python ./datasets/visualize_vessel_graphs.py --source_dir ./datasets/dataset_18_June_2023 --out_dir ./datasets/dataset_18_June_2023_png --factor 4

# Create label maps
python ./datasets/visualize_vessel_graphs.py --source_dir ./datasets/synth_csv --out_dir ./datasets/dataset_18_June_2023_label_png --factor 16 --binarize

To turn them into realistic images, transform the images using a pretrained GAN. Update the source path in the model config file and run:

python test.py --config_file [PATH_TO_CONFIG_FILE] --epoch [EPOCH]