flare_21

The solution for this challenge is based on the nnUNet repository. Give thanks to Fabian’s great work!

We made three major modifications on top of the original nnUNet:

1. In the inference stage, we only used the two samples with the largest flipping difference as test time augmentation (TTA).
2. In the deep supervision block, we only used the decoder output of the last three layers for weighted average.
3. To reduce performance degradation caused by network simplification, we referred to the idea of OHEM and used Top-k loss in cross-entropy to supervise the voxels with the largest difference, which would help maintain the performance as much as possible under the simplified structure.

Update 2021.10.23

I have retrained and uploaded the docker file based on the Abdomen1KCT-fullySupervised dataset, which can be obtained through Docker Hub or Google Netdisk. The installation method can refer to the docker chapter below but change the instructions to:

# (Optional) download from Docker Hub
docker pull ttime/abdomen1k
#subtask 1
docker container run --gpus "all" --name ttime_abdomen1k --rm -v $PWD/inputs/:/workspace/inputs/ -v $PWD/outputs/:/workspace/outputs/ ttime_abdomen1k:latest /bin/bash -c "sh predict.sh 1"
#subtask 2
docker container run --gpus "all" --name ttime_abdomen1k --rm -v $PWD/inputs/:/workspace/inputs/ -v $PWD/outputs/:/workspace/outputs/ ttime_abdomen1k:latest /bin/bash -c "sh predict.sh 2"

Installation

from source (recommend)

1. Clone the directory and its submodule (nnUNet):

   git clone --recursive https://github.com/tea321000/flare21_solution
   cd flare21_solution
   # if you are updating an existing checkout
   git submodule sync
   git submodule update --init --recursive --jobs 0

2. Follow the installation instructions of my forked repository to install the modified version of nnUNet. Ensure that the relevant environment variables have been set similar to the following commands:

   # some APIs in the recently updated version of batchgenerators have changed, so it is recommended to use the old version
   pip install batchgenerators==0.21
   cd nnUNet
   pip install -e .
   # you can write environment variables into .bashrc
   export nnUNet_raw_data_base="/media/fabian/nnUNet_raw"
   export nnUNet_preprocessed="/media/fabian/nnUNet_preprocessed"
   export RESULTS_FOLDER="/media/fabian/nnUNet_trained_models"
   # check if the environment variable has been set
   echo $nnUNet_raw_data_base
   echo $nnUNet_preprocessed
   echo $RESULTS_FOLDER

docker

We also provide docker images for the competition, but this method of installation is not convenient for modification, so it is recommended to use it only for inference. You can use it with the following instructions:

docker image load < ttime_flare_final.tar.gz
docker container run --gpus "all" --name ttime --rm -v $PWD/inputs/:/workspace/inputs/ -v $PWD/outputs/:/workspace/outputs/ teamname:latest /bin/bash -c "sh predict.sh"

Training

At present, the 2d network has not been completely modified yet, so my solution can only be used to train the 3d network. If you have any ideas or questions, you can post them in the issue or send an email to me. The training instructions of the 3d network is the same as the original nnUNet, but the Trainer needs to be changed to nnUNetTrainerV2_flare:

# my solution in challenge uses lowres network, but fullres network can also be used
nnUNet_train 3d_lowres nnUNetTrainerV2_flare TaskXXX_MYTASK FOLD --npz

Inference

1. Download the trained model from the download.sh script or the Google Drive link(task id is 500):

   bash download.sh

   If you download from Google Drive link, you need to put the compressed file into the manually created trained_weight folder in flare21_solution folder and run download.sh.

2. Create inputs folder under flare21_solution , put the samples to be predicted into the folder, and run predict.sh:

   bash predict.sh

The predictions are saved in the outputs folder by default, and the input and output paths can be changed by modifying the script.