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.
- 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 bypip install -e .
- Set
paths.py
Please strictly follow this path setting guidance unless you know what you are changing!
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
)
- imagesTr (file names shoud be
Then, Put the above Task07_iSeg
folder into path to/mydata_folder/nnUNet_raw_splitted
- 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
- 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
- 3D U-Net is a strong baseline!
- The performance has significantly drop on new sites dataset.
I use all the training cases during training.
- Put the
dataset.json
intopath 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 inpath to nnUNet/nnunet
(the same path topaths.py
file).