SegRap2023

1. Totorial for Algorithm Docker Image (official guideline of grand challenge)

1.1 Important input and output

• For task1, the input dir is /input/images/head-neck-ct/ (non-contrast-ct images) and /input/images/head-neck-contrast-enhanced-ct/ (contrast-ct images). The output dir is /output/images/head-neck-segmentation/. Note that the final prediction has to be a 4D .mha file, which array shape is [45, *image_shape]. An example code and output is shown as Docker_tutorial/stacked_results_to_4d_mha.py and Docker_tutorial/oars_output_example.mha.

• For task2, the dir is /input/images/head-neck-ct/ (non-contrast-ct images) and /input/images/head-neck-contrast-enhanced-ct/ (contrast-ct images). The output dir is /output/images/gross-tumor-volume-segmentation/. Note that the final prediction has to be a 4D .mha file, which array shape is [2, *image_shape]. An example code and output is shown as Docker_tutorial/stacked_results_to_4d_mha.py and Docker_tutorial/gtvs_output_example.mha.

1.2 Algorithm examples based on nnUNet

We provide two algorithm examples based on nnUNet, which is only the baseline for two tasks. If your method is based on nnUNet, you can follow the example to generate predictions and run sh export.sh to generate an Algorithm Container Image in tar.gz format. The details about loading input, generating predictions, and saving output can be seen in the process.py.

In addition, you can download the example data and model weight from GoogleDrive and BaiduNetDisk to the folder images and weight, respectively. Before submitting, you can test the docker image on your local machine by running sh test.sh or sudo sh test.sh, we show an example output on our ubuntu22.04 (one 3090 GPU).

1.2.1 How to test the container locally?

1. Parepare your images and weight as following format

   • SegRap2023_task1_OARs_nnUNet_Example
     • images
       • images
         • head-neck-contrast-enhanced-ct
           • segrap_0001.mha
         • head-neck-ct
           • segrap_0001.mha
     • weight
       • fold_0
         • model_final_checkpoint.model
         • model_final_checkpoint.model.pkl
         • plans.pkl

2. How about the output?

   • You can check out if there are predictions in the output folder /output/images/head-neck-segmentation or /output/images/gross-tumor-volume-segmentation that are corresponded to the input images. Run the following command will show the files in the output folder.

     docker run --rm \
         -v segrap2023_segmentationcontainer-output-$VOLUME_SUFFIX:/output/ \
         python:3.10-slim ls -al /output/images/head-neck-segmentation

   • The test folder is just an empty folder which hasn't been used in the docker image, so you can ignore or remove it.

   • You can ignore the error No such file or directory: '/output/results.json' when you run the docker locally.

1.3 Algorithm examples based on others.

If your method is not based on nnUNet, you can modify the function of predict() in process.py and other corresponding parts for inference. It's easy to read and modify, but please ensure the format of the output file (a 4D .mha, the right mapping between the index of 4D file and OARs or GTVs.). We provided an example (Docker_tutorial/stacked_results_to_4d_mha.py) to stack individual oars/gtvs predictions of a patient into a required 4d .mha files.

1.4 Q&A.

If you meet any questions when submitting your docker images, you can email us (luoxd1996@gmail.com or fujia98914@gmail.com), or post the issue or discuss it in the forum at any time.

2. How to submit the algorithm?

1. If you have not created your algorithm, you can go to https://segrap2023.grand-challenge.org/evaluation/challenge/algorithms/create/ to create an algorithm with 30GB memory.

2. Upload your Algorithm Container Image, then wait for the container to be active.

3. Go to the SegRap2023 submit website, choose the task and submit your Algorithm Image.

4. After submitting, you can wait for the update of Leaderboards.

3. Tutorial for SegRap2023 Challenge

This repository provides tutorial code for Segmentation of Organs-at-Risk and Gross Tumor Volume of NPC for Radiotherapy Planning (SegRap2023) Challenge. Our code is based on PyMIC, a pytorch-based toolkit for medical image computing with deep learning, that is lightweight and easy to use.

Requirements

This code depends on Pytorch, PyMIC. To install PyMIC and GeodisTK, run:

pip install PYMIC

Segmentation model based on PyMIC

Dataset and Preprocessing

• Download the dataset from SegRap2023 and put the dataset in the data_dir/raw_data.

• For data preprocessing, run:

  python Tutorial/preprocessing.py

  This will crop the images with the maximal nonzero bounding box, and the cropped results are normalized based on the intensity properties of all training images. By setting the args.task to OARs and GTVs, we can get the preprocessed images and labels for two tasks that are saved in data_dir/Task001_OARs_preprocess and data_dir/Task002_GTVs_preprocess, respectively.

Training

• Run the following command to create csv files for training, validation, and testing. The csv files will be saved to config/data_OARs and config/data_GTVs.

  python Tutorial/write_csv_files.py

• Run the following command for training and validation. The segmentation model will be saved in model/unet3d_OARs and model/unet3d_GTVs, respectively.

  pymic_train Tutorial/config/unet3d_OARs.cfg
  pymic_train Tutorial/config/unet3d_GTVs.cfg

  Note that you can modify the settings in .cfg file to get better segmentation results, such as RandomCrop_output_size, loss_class_weight, etc.

Testing

• After training, run the following command, we can get the performance on the testing set, and the predictions of testing data will be saved in result/unet3d_OARs and result/unet3d_GTVs.

  pymic_test Tutorial/config/unet3d_OARs.cfg
  pymic_test Tutorial/config/unet3d_GTVs.cfg

Postprocessing

• Run the following command to obtain the final predictions, which are saved in result/unet3d_OARs_post and result/unet3d_GTVs_post.

  python Tutorial/postprocessing.py

Segmentation model based on nnUNet

Postprocessing

• Following Tutorial/nnunet_baseline.ipynb, you can obtain the final predictions based on the outputs from nnUNet. In addition, you also can use Tutorial/preprocessing.py for preprocessing firstly (details as mentioned in the above tutorial) and then train networks using Tutorial/nnunet_baseline.ipynb.

4. Evaluation for SegRap2023 Challenge

Run following command to get the quantitative evaluation results.

python Eval/SegRap_Task001_DSC_NSD_Eval.py
python Eval/SegRap_Task002_DSC_NSD_Eval.py

Reference

@article{luo2023segrap2023,
  title={SegRap2023: A Benchmark of Organs-at-Risk and Gross Tumor Volume Segmentation for Radiotherapy Planning of Nasopharyngeal Carcinoma},
  author={Luo, Xiangde and Fu, Jia and Zhong, Yunxin and Liu, Shuolin and Han, Bing and Astaraki, Mehdi and Bendazzoli, Simone and Toma-Dasu, Iuliana and Ye, Yiwen and Chen, Ziyang and others},
  journal={arXiv preprint arXiv:2312.09576},
  year={2023}
}

@article{wang2023pymic,
  title={PyMIC: A deep learning toolkit for annotation-efficient medical image segmentation},
  author={Wang, Guotai and Luo, Xiangde and Gu, Ran and Yang, Shuojue and Qu, Yijie and Zhai, Shuwei and Zhao, Qianfei and Li, Kang and Zhang, Shaoting},
  journal={Computer Methods and Programs in Biomedicine},
  volume={231},
  pages={107398},
  year={2023},
  publisher={Elsevier}
}