This is the official pytorch implementation of Paper "Topology-Preserving Hard Pixel Mining for Tubular Structure Segmentation", BMVC 2023. [PDF] [Video] [Supplementary]
Install Requirements:
conda install --yes --file requirements.txt
Install topohpm:
python setup.py install
Given the package topohpm installed, we can train the network through following command:
train3dunet --config <CONFIG>
To train the network, we need to specify a config file. We have provide several config files for different datasets and different losses in resource folder. In our experiments, the method is evaluated through 5-fold cross validation. We use the code topohpm/scripts/random_split_k_fold.py to randomly split the dataset into k fold.
python random_split_k_fold.py -r <raw_dir> -l <label_dir> -o <output_dir> -k <kfold=5>
train.sh is the script for k-fold batch training:
nohup ./train.sh &
Given the package topohpm installed, we can run the prediction via:
predict3dunet --config <CONFIG>
Similar to the training process, we need to specify a config file for prediction. predict.sh is the script for k-fold batch prediction, which saves all the predictions into generated_files/predictions:
nohup ./predict.sh &
We compute Dice coefficient and MeanIoU to evaluate the volumetric accuracy, Variation of Information (VOI) and Adapted Rand Error (ARE) for clustering-based metrics, Betti-0 and Betti-1 error as topology measurements, respectively. Evaluation on 2D image and 3D image can be done through topohpm/scripts/evaluate.py and topohpm/scripts/evaluate3d.py, respectively.
@inproceedings{Zhang_2023_BMVC,
author = {Guoqing Zhang and Caixia Dong and Yang Li},
title = {Topology-Preserving Hard Pixel Mining for Tubular Structure Segmentation},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0846.pdf}
}
The code structure is mainly inherited from pytorch-3dunet.