LiTS-liver-and-lesion-segmentation-toolbox

This Toolbox is designed for the research purpose only. Our research report can be found here.

Environment setup

We have developed our Toolbox with pytorch under python=3.6 and tested for both (CUDA 10.0 + pytorch=1.3.0) and (CUDA 10.2 + pytorch=1.0.0).

Build virtual environment:

  $ conda create --name <envname> --file requirements.txt

If you want to install packages manually, here is the list (pip install -)

• medpy (read medical image)
• nibabel (read medical image)
• tqdm
• opencv-python
• h5py
• pytorch, torchvision (please see the offical site for installation details)
• torchsummary

Usage

Under the root folder:

• Put the unzipped dataset downloaded from the site to the folder ./data. Our pre-processing can find the *.nii file recursively.

Interactive jupyter widgets for dataset visualization

• run the jupyter notebook: visualize_imgs.ipynb

Data pre-processing

• The train/valid set will be saved into two hdf5 files separately with name train_LiTS_db.h5 and valid_LiTS_db.h5

  $ python ./src/preprocessing/preprocessing.py -dir "./data" --save2file "./data/LiTS_db.h5" --valid-split 0.2
  

• If you want to resize the volumes (per z-slice)

  $ python ./src/preprocessing/preprocessing.py -dir "./data" --save2file "./data/LiTS_db_224.h5" --img-shape "(224,224)"
  

• For the test set (without segmentation mask): the file name will be test_LiTS_db.h5

  $ python ./src/preprocessing/preprocessing.py -dir "./data" --save2file "./data/LiTS_db.h5" --test-set
  

Interactively visualize dataloader (with data augmentation)

  $ python ./src/preprocessing/dataloader.py -f ./data/train_LiTS_db_224.h5 --shuffle

Training

• From scratch:

  $ python ./src/train.py --epochs 50 -lr 0.00005 --log-dir "./weights/Exp_000" -trainf "./data/train_LiTS_db.h5" -validf "./data/valid_LiTS_db.h5" --batch-size 32  --num-cpu 32 --shuffle --cuda 0
  

• Load pretrained weights or model

  $ python ./src/train.py --epochs 50 -lr 0.00005 --log-dir "./weights/Exp_000" -trainf "./data/train_LiTS_db.h5" -validf "./data/valid_LiTS_db.h5" --batch-size 32  --num-cpu 32 --shuffle --cuda 0 --load-weights "./weights/model_weights.pth"
  

Inference and evaluation

Note that we can either load pretrained weights (model.state_dict()) by --load-weights or the entire model by --load-model

• Evaluation mode

  $ python ./src/inference.py **-eval** --batch-size 32 --num-cpu 32 -testf ./data/valid_LiTS_db_224.h5 --load-weights "./weights/Exp_000/model_weights.pth" --save2dir ./results/Exp_000
  

• Inference mode

  $ python ./src/inference.py --batch-size 32 --num-cpu 32 -testf ./data/test_LiTS_db_224.h5 --load-model "./weights/Exp_000/model.pth" --save2dir ./results/Exp_000
  

Submission results

---

TODO list

• [dataset] medpy dataset header manipulation
• [dataset] file format selection
• [data augmentation] compatibility between opencv and pytorch's dataloader: muliti-processing issue with numpy.random in Linux
• [data augmentation] separate geometry transform and pixel-wise transform
• [preprocessing] resize/transform volumetric mask
• [dataloader] hdf5 file compatibility with pytorch's muliti-processing
• Statistic of dataset
• [device] enable Multi-GPU training (with NVLink): data vs. model parallelism in pytorch
• [monitoring] Add pytorch's tensorboard to ./src/train.py
• Data augmentation zoo
• loss functions zoo
• model zoo
• Hyperparameters tuning module
• Add domain adaptation module