Automatic Radiotherapy Treatment Planning ， Knowledge-Based Planning ， Dose Prediction ， Cascade 3D Network (C3D) ，DCNN, Head and Neck , 1st Place Solution to the AAPM OpenKBP challenge

Please feel free to concat me if you have any questions， email: [email protected], Shuolin Liu

This repository contains an PyTorch implementation for radiotherapy dose prediction, along with pre-trained models and examples.

The goal of this implementation is to be simple, highly extensible, and easy to integrate into your own projects. This implementation is a work in progress -- more dose prediction models are currently being implemented. Currently support：

• C3D (under review), a cascade 3D network for radiotherapy dose prediction, the 1st place solution to the AAPM OpenKBP challenge
  (Official OpenKBP paper is now available on arXiv)

• DCNN, a lightweight and accurate dose prediction method

• Results on OpenKBP Test Set using a Single model with test-time augmenation(TTA)

• OpenKBP leaderboard

• torch >=1.2.0
• tqdm
• opencv-python
• numpy
• SimpleITK
• pandas
• scikit-image
• scipy

1. Data Preparation

   • Download OpenKBP challenge repository, and copy the repository to
     /path_to_your_RTDosePrediction/RTDosePrediction/Data/

     For me, /path_to_your_RTDosePrediction/ is E://Project/RTDosePrediction-main/

   • C3D：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/DataPrepare
     python prepare_OpenKBP_C3D.py
     

     The training Data will be saved in /path_to_your_RTDosePrediction/RTDosePrediction/Data/OpenKBP_C3D

   • DCNN：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/DataPrepare
     python prepare_OpenKBP_DCNN.py
     

     The training Data will be saved in /path_to_your_RTDosePrediction/RTDosePrediction/Data/OpenKBP_DCNN

2. Training

   • C3D：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/C3D
     python train.py --batch_size 2 --list_GPU_ids 1 0 --max_iter 80000
     

     Larger batch_size will bring more stable results. If you want to train C3D with batch size of 4, use:
     

     python train.py --batch_size 4 --list_GPU_ids 3 2 1 0 --max_iter 80000

   • DCNN：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/DCNN
     python train.py --batch_size 32 --list_GPU_ids 0 --max_iter 100000
     

3. Testing

   The prediction results will be saved in /path_to_your_RTDosePrediction/RTDosePrediction/Output/XXX/Prediction

   • C3D：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/C3D
     python test.py --GPU_id 0 
     

   • DCNN：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/DCNN
     python test.py --GPU_id 0 
     

4. Using pre-trained models

   • Download model weights for C3D (Google Drive, Baidu Drive, PassWord：voni) and DCNN(Google Drive, Baidu Drive, PassWord：j56y)

   • Copy model weights to /path_to_your_RTDosePrediction/RTDosePrediction/PretrainedModels

   • C3D：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/C3D
     python test.py --GPU_id 0 --model_path ../../PretrainedModels/C3D_bs4_iter80000.pkl
     

   • DCNN：

     cd /path_to_your_RTDosePrediction/RTDosePrediction/Src/DCNN
     python test.py --GPU_id 0 --model_path ../../PretrainedModels/DCNN_bs32_iter100000.pkl
     

if you find C3D and DCNN useful in your research, please consider citing:

• C3D

@article{C3D,
   title = {Cascade 3D Network for Radiotherapy Dose Prediction : 1st Place Solution to OpenKBP Challenge},
   author = {Shuolin Liu and Jingjing Zhang and Teng Li and Hui Yan  and Jianfei Liu},
   journal = {Medical Physics, under review}
}

• DCNN

@article{DCNN,
   title = {Predicting voxel-level dose distributions for esophageal radiotherapy using densely connected network with dilated convolutions},
	doi = {10.1088/1361-6560/aba87b},
	url = {https://doi.org/10.1088%2F1361-6560%2Faba87b},
	year = 2020,
	month = {oct},<br>
	publisher = {{IOP} Publishing},
	volume = {65},
	number = {20},
	pages = {205013},
	author = {Jingjing Zhang and Shuolin Liu and Hui Yan and Teng Li and Ronghu Mao and Jianfei Liu},
	journal = {Physics in Medicine {\&} Biology
}

Thank OpenKBP Organizers： Aaron Babier, Binghao Zhang, Rafid Mahmood, Timothy Chan， Andrea McNiven， Thomas Purdie， and Kevin Moore.

• https://github.com/ababier/open-kbp