/FLARE23

Primary LanguagePython

Environments and Requirements

  • Ubuntu 20.04 LTS
  • RTX 3090 with 24GB GPU memory
  • 11.7
  • Python 3.10

To install requirements:

git clone https://github.com/luckieucas/FLARE23.git
cd nnUNet
pip install -e .

Dataset

  • MICCAI FLARE 2023 dataset

Preprocessing

  • cropping
  • intensity normalization
  • resampling
  • flip
  • rotation
  • scale

Running the data preprocessing code:

nnUNetv2_plan_and_preprocess -d 12 --verify_dataset_integrity

Training

  1. To train the model(s) in the paper, run this command:
python run_training_Flare.py 12 3d_mylowres 1 -tr nnUNetTrainerFlarePseudoCutUnsupLow -p nnUNetPlans

Inference

  1. To infer the testing cases, run this command:
nnUNetv2_predict -i <path_to_data> -o  <path_to_output_data>  -d 12 -c 3d_mylowres -f 1 -chk <name_of_trained_model> -tr  nnUNetTrainerFlarePseudoCutUnsupLow -step_size 0.6 -npp 3 --disable_tta

Evaluation

To compute the evaluation metrics, run:

nnUNetv2_evaluate_folder <path_to_ground_truth>  <path_to_inference_results>

Results

Our method achieves the following performance on MICCAI FLARE23: Fast, Low-resource, and Accurate oRgan and Pan-cancer sEgmentation in Abdomen CT

Model name Organ DICE(%) Organ NSD(%) Tumor DICE(%) Tumor NSD(%)
Our method 92.18 96.33 46.26 38.65

Contributing

Pick a license and describe how to contribute to your code repository.

Acknowledgement

We thank the contributors of public datasets.