JunMa11 / iseg2019

Solution to MICCAI iSeg-2019

iSeg2019 Homepage & Leaderboard

3D U-net Baseline for iseg-2019 and An Attempt at Dealing With Multiple Sites Data via Histogram Matching

The 3D U-Net baseline is based on nnU-Net.

Requirement

• Pytorch version >=1.0.1
• nnU-Net
  • git clone https://github.com/MIC-DKFZ/nnUNet.git
  • cd nnUNet
  • Install with pip install -r requirements.txt followed by pip install -e .
• Set paths.py
  • mkdir mydata_folder (mydata_folder should be located in the same path with paths.py)
  • line 35: base = "path to/mydata_folder"
  • line 51: preprocessing_output_dir = "path to/mydata_folder/pre_data"
  • line 60: network_training_output_dir = os.path.join(base, my_output_identifier)

Please strictly follow this path setting guidance unless you know what you are changing!

Preprocessing: convert and rename hdr files to nii files by hdr2nii

The finally folder structure should be

• Task07_iSeg
  • imagesTr (file names shoud be iseg_1_0000.nii.gz, iseg_1_0001.nii.gz, ..., iseg_10_0000.nii.gz, iseg_10_0001.nii.gz)
  • labelsTr (file names shoud be iseg_1.nii.gz, ..., iseg_10.nii.gz)
  • imagesVal (file names shoud be iseg_11_0000.nii.gz, iseg_11_0001.nii.gz, ..., iseg_23_0000.nii.gz, iseg_23_0001.nii.gz)
  • imagesTs (file names shoud be iseg_24_0000.nii.gz, iseg_24_0001.nii.gz, ..., iseg_39_0000.nii.gz, iseg_39_0001.nii.gz)

Then, Put the above Task07_iSeg folder into path to/mydata_folder/nnUNet_raw_splitted

Testing

• Downloda the pre-trained model and put it in the nnU-Net's model folder mydata_folder/nnUNet/3d_fullres/Task07_iSeg. Download: BaiduNetDisk pw:xyhu
• Inference validation set: Run python inference/predict_simple.py -i path to/imagesVal -o OUTPUT_FOLDER -t Task07_iSeg -tr nnUNetTrainer -m 3d_fullres -f all
• Inference testing set: Run python inference/predict_simple.py -i path to/imagesTs -o OUTPUT_FOLDER -t Task07_iSeg -tr nnUNetTrainer -m 3d_fullres -f all

Histogram matching

• Generate histogram matching results by running HistogramMatch.py
• Inference histogram matching results: Run python inference/predict_simple.py -i path to/imagesTsHist -o OUTPUT_FOLDER -t Task07_iSeg -tr nnUNetTrainer -m 3d_fullres -f all

Results

• 3D U-Net is a strong baseline!
• The performance has significantly drop on new sites dataset.

Re-train the model on TitanXP GPU

I use all the training cases during training.

• Put the dataset.json into path to/mydata_folder/nnUNet_raw_splitted/Task07_iSeg
• Run python experiment_planning/plan_and_preprocess_task.py -t Task07_iSeg -pf 10
• Run python run/run_training.py 3d_fullres nnUNetTrainer Task07_iSeg all --ndet.

All the python command should be run in path to nnUNet/nnunet (the same path to paths.py file).