/RTDosePrediction

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

Primary LanguagePythonMIT LicenseMIT

RTDosePrediction

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: 1980073622@qq.com, Shuolin Liu

Warning， this repository will be closed !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Due to some conflicts of interest, this repository will be closed in two weeks. If you wish to continue using it, please copy the codes and save them to your own machine. The licence for these codes is still the MIT licence.

Overview

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: a cascade 3D network for radiotherapy dose prediction, the 1st place solution to the AAPM OpenKBP challenge

  • DCNN: a lightweight dose prediction method

  • HD-Unet: a hierarchically densely connected U-net, Dan Nguyen et al. Compared to original HD-Unet paper, we add instance normalization (IN) for better performance. We reported the results of original HD-Unet in C3D paper

Performance

  • Results on OpenKBP Test Set using a Single model with test-time augmenation(TTA)
Model Batch
size		 GPU
memory		 Training
iterations		 Training
time		 Dose
score		 DVH
score		 Pre-trained
Models
C3D
(3D)		 2 18Gb 80,000 50 hours
(Two 1080TIs)		 2.46 1.46 Google Drive
Baidu Drive, PassWord：voni
DCNN
(2D)		 32 3Gb 100,000 20 hours
(Single 1080TI)		 2.75 1.68 Google Drive
Baidu Drive, PassWord：j56y
HD-Unet
(3D)		 2 16Gb 80,000 - hours
(Single A100)		 2.59 1.64 Google Drive
Baidu Drive, PassWord：cc61
  • OpenKBP leaderboard

Requirements

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

Usage

  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

Citation

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

  • C3D
@article{C3D,
   title = {Technical Note: A Cascade 3D U-Net for Dose Prediction in Radiotherapy},
   author = {Shuolin Liu and Jingjing Zhang and Teng Li and Hui Yan  and Jianfei Liu},
   journal = {Medical Physics,  https://doi.org/10.1002/mp.15034}
}
  • 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
}

Acknowledgement

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