Reinforced Active Learning for Medical Image Segmentation (RALMIS)

Code (https://github.com/ArantxaCasanova/ralis) from the paper [Reinforced Active Learning for Image Segmentation] (https://arxiv.org/abs/2002.06583) adapted to 3D medical images.

Dependencies

• python 3.9
• numpy 1.20.2
• scipy 1.6.2
• Pytorch 1.8.1
• Torchvision 0.9.1
• PIL 8.2.0

Scripts

The folder 'scripts' contains the different bash scripts that could be used to train the same models used in the master thesis [Deep Reinforcement Learning for Interactive Training of Medical Image Segmentation Networks], for both anatomical cardiac (ACDC) and brain tumour (BraTS2018) segmentation datasets.

• Segmentation networks' backbone pre-trained on ImageNet
  • a) on the ACDC cardiac dataset:
    • exp2_final_acdc_baselines_*.sh to train the baseline active learning methods 'random', 'entropy' and 'bald' and train/test the segmentation network
    • exp2_final_acdc_ralis_train.sh to train the reinforcement agent for the 'ralis' model
    • exp2_final_acdc_ralis_test.sh to test the reinforcement agent for the 'ralis' model and train/test the segmentation network
  • b) on the BraTS2018 brain tumour dataset:
    • exp2b_brats_baselines_*.sh to train the baseline active learning methods 'random', 'entropy' and 'bald' and train/test the segmentation network
    • exp2b_brats18_train.sh to train the reinforcement agent for the 'ralis' model
    • exp2b_brats18_test.sh to test the reinforcement agent for the 'ralis' model and train/test the segmentation network

Furthermore the folder 'scripts' contains the bash scripts that could be used to train the same models used in the paper from Casanova et al., for both Camvid and Cityscapes datasets.

• launch_supervised.sh: To train the pretrained segmentation models.
• launch_baseline.sh: To train the baselines 'random', 'entropy' and 'bald'.
• launch_train_ralis.sh: To train the 'ralis' model.
• launch_test_ralis.sh: To test the 'ralis' model.

Datasets

Our investigated datasets:

• Anatomical cardiac structure segmentation dataset [ACDC]: https://www.creatis.insa-lyon.fr/Challenge/acdc/index.html
• Pathological brain tumour segmentation dataset [BraTS18]: https://www.med.upenn.edu/sbia/brats2018/data.html

Trained models

To download the trained reinforcement learning agent on the ACDC and BraTS datasets: https://drive.google.com/drive/folders/1SkBdh5HVZsO8Og8dgFSIVKvYiCASCGIY?usp=sharing

Notes by C Baumgartner for use in Tuebingen ML Cloud

Additions to base code:

• add number of slurm job files to scripts
• add a template for running an interactive training to scripts
• make code_path an additional argument to prevent hard coded img_paths
• remove dataset from all hardcoded path variables in dataset classes

TODOs to make code run

• Change paths in utils/parser.py
• Change paths in scripts/slurm_* and scripts/interative_slurm_call_debug.sh to match system

Run code on Tue ML Coud Slurm

To run ralis training execute

sbatch devel/ralis/scripts/slurm_train_ralis.sh

To run stuff depending on the pre-trained models, don't forget to download them from the Google Drive link below and copy into your checkpoints folder (e.g. ckpt_seg)

Citation

If you use this code, please cite the original paper:

@inproceedings{
Casanova2020Reinforced,
title={Reinforced active learning for image segmentation},
author={Arantxa Casanova and Pedro O. Pinheiro and Negar Rostamzadeh and Christopher J. Pal},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=SkgC6TNFvr}
}