Continuous U-Net: Faster, Greater and Noiseless

This repository contains the code to reproduce the experiments from Continuous U-Net: Faster, Greater and Noiseless.

Dataset

The dataset used in the paper can be download at relevant links below, please place (rename and unzip) the downloaded item directly at corresponding folders.

We put it under the folder ./datasets/.

./datasets
├── Cells
├── STARE 
    └── BloodVesselSegmentation
        ├── inputs
        └── GT
├── nuclei/train  *require Kaggle signin
├── breast_cancer/Dataset_BUSI_with_GT  *require Kaggle signin
├── Kvasir-SEG
└── melanoma_segmentation
    ├── train
        ├── ISBI2016_ISIC_Part1_Training_Data
        └── ISBI2016_ISIC_Part1_Training_GroundTruth
    └── test
        ├── ISBI2016_ISIC_Part1_Test_Data
        └── ISBI2016_ISIC_Part1_Test_GroundTruth

The outputs from the trainning will be saved under folder ./outputs/.

Get started

Clone current repository:

git clone https://github.com/Math-ML-X/cUNet-private

Download datasets with links in Dataset.

Create virtual environment with Conda using ./usode.yml:

conda env create -f usode.yml

Activate the virtual environment:

conda activate usode

Training with './script.py':

python script.py --dataset STARE --net USODE --solver rk4

(default setting on STARE dataset with our method and the Runge–Kutta method (RK4) for solver).

Cite

If this repository is useful for your research, please cite:

@article{cheng2023continuous,
  title={Continuous U-Net: Faster, Greater and Noiseless},
  author={Cheng, Chun-Wun and Runkel, Christina and Liu, Lihao and Chan, Raymond H and Sch{\"o}nlieb, Carola-Bibiane and Aviles-Rivero, Angelica I},
  journal={arXiv preprint arXiv:2302.00626},
  year={2023}
}

Acknowledgments

We thank @YanqiCheng for code clean and refactoring, which makes it well structured and easy to understand.