Skip to content

scil-vital/TrackToLearn

Repository files navigation

Track-to-Learn/TractOracle-RL: reinforcement learning for tractography.

TractOracle-RL is half of TractOracle, a reinforcement learning system for tractography. TractOracle-RL is a tractography algorithm which is trained via reinforcement learning using TractOracle-Net.

See Versions for past and current interations.

Getting started

Right now, only python 3.10 is supported.

It is recommended to use a python virtual environment to run the code.

virtualenv .env --python=python3.10
source .env/bin/activate

Then, install the dependencies and setup the repo with

./install.sh

Getting errors during installation ? Open an issue !

Tracking

You will need a trained agent for tracking. One is provided in the models folder and is loaded autmatically when tracking. You can then track by running ttl_track.py.

usage: ttl_track.py [-h] [--input_wm] [--sh_basis {descoteaux07,tournier07}]
                    [--compress thresh] [-f] [--save_seeds] [--agent AGENT]
                    [--hyperparameters HYPERPARAMETERS] [--n_actor N]
                    [--npv NPV] [--min_length m] [--max_length M]
                    [--noise sigma] [--fa_map FA_MAP]
                    [--binary_stopping_threshold BINARY_STOPPING_THRESHOLD]
                    [--rng_seed RNG_SEED]
                    in_odf in_seed in_mask out_tractogram

 Generate a tractogram from a trained model. See `--help` for usage.

You will need to provide fODFs, a seeding mask and a WM mask. The seeding mask must represent the interface of white matter and gray matter. WM tracking is no longer supported.

Agents used for tracking are constrained by their training regime. For example, the agents provided in models were trained on a volume with a resolution of ~1mm iso voxels and a step size of 0.75mm using fODFs of order 8, descoteaux07 basis. When tracking on arbitrary data, the step-size and fODF order and basis will be adjusted accordingly automatically (i.e resulting in a step size of 0.375mm on 0.5mm iso diffusion data). However, if using fODFs in the tournier07 basis, you will need to set the --sh_basis argument accordingly.

Docker

You can also track using Docker. First pull the container by running

sudo docker pull scilus/tractoracle-rl:v2024b

You can then track by running

sudo docker run scilus/tractoracle-rl:v2024b ttl_track.py ...

See Docker volumes to use local files. To use CUDA capabilities with Docker, you will need to install the NVIDIA Container Toolkit. You will then be able to use the --gpus flag. For example:

sudo docker run --gpus all scilus/tractoracle-rl:v2024b ttl_track.py ...

Training

First, make a dataset .hdf5 file with TrackToLearn/dataset/create_dataset.py.

usage: create_dataset.py [-h] [--sh_basis {descoteaux07,tournier07}]
                         config_file output

positional arguments:
  config_file  Configuration file to load subjects and their volumes.
  output       Output filename including path

optional arguments:
  -h, --help   show this help message and exit

Basis options:
  --sh_basis {descoteaux07,tournier07} Spherical harmonics basis used for the SH coefficients.

Example dataset config files are available in examples.

Then, you may train an agent by running python TrackToLearn/trainers/sac_auto_train.py.

usage: sac_auto_train.py [-h] [--workspace WORKSPACE] [--rng_seed RNG_SEED]
                         [--use_comet] [--n_dirs N_DIRS]
                         [--binary_stopping_threshold BINARY_STOPPING_THRESHOLD]
                         [--n_actor N_ACTOR] [--hidden_dims HIDDEN_DIMS]
                         [--max_ep MAX_EP]
                         [--log_interval LOG_INTERVAL] [--lr LR] [--gamma GAMMA]
                         [--alignment_weighting ALIGNMENT_WEIGHTING] [--npv NPV]
                         [--theta THETA] [--min_length m] [--max_length M]
                         [--step_size STEP_SIZE] [--noise sigma]
                         [--oracle_checkpoint ORACLE_CHECKPOINT]
                         [--oracle_validator] [--oracle_stopping_criterion]
                         [--oracle_bonus ORACLE_BONUS]
                         [--scoring_data SCORING_DATA]
                         [--tractometer_reference TRACTOMETER_REFERENCE]
                         [--tractometer_validator]
                         [--tractometer_dilate TRACTOMETER_DILATE]
                         [--alpha ALPHA] [--batch_size BATCH_SIZE]
                         [--replay_size REPLAY_SIZE]
                         path experiment id dataset_file

To use Comet.ml, follow instructions here, with the config file either in your home folder or current folder. Usage of comet-ml is necessary for hyperparameter search. This constraint may be removed in future releases.

Contributing

Contributions are welcome ! There are several TODOs sprinkled through the project which may inspire you. A lot of the code's architecure could be improved, better organized, split and reworked to make the code cleaner. Several performance improvements could also easily be added.

Versions

TractOracle-RL (2024b)

Théberge, A., Descoteaux, M., & Jodoin, P. M. (2024). TractOracle: towards an anatomically-informed reward function for RL-based tractography. Submitted to MICCAI 2024.

The reference commit to the main brain for this work is 0fb20306edc32b6015fbfe9b79677015cd0602cf. Please use this commit as starting point if you want to build upon Track-to-Learn (TractOracle-RL). See README above for usage.

See preprint: https://arxiv.org/pdf/2403.17845.pdf

See additional data: https://zenodo.org/doi/10.5281/zenodo.10931313

Conference paper (hopefully) coming soon.

What matters in Reinforcement Learning for Tractography (2024a)

Théberge, A., Desrosiers, C., Boré, A., Descoteaux, M., & Jodoin, P. M. (2024). What matters in reinforcement learning for tractography. Medical Image Analysis, 93, 103085.

The reference commit to the main branch for this work is 9f97eefbbdb05a2c90ea74e8384ac2891b194a3e. Please use this commit as starting point if you want to reproduce or build upon the work of the 2024a paper.

See journal paper: https://www.sciencedirect.com/science/article/pii/S1361841524000100

See preprint: https://arxiv.org/abs/2305.09041

Incorporating anatomical priors into Track-to-Learn (2022)

The reference commit to the main branch for this work is dbae9305b4a3e9f21c3249121ef5dc5ed9faa899.

This work is presented at the ISMRM Workshop on Diffusion MRI: From Research to Clinic, poster #34 (email me for the abstract and poster). This work adds the use of Continuous Map Criterion (CMC, https://www.sciencedirect.com/science/article/pii/S1053811914003541), asymmetric fODFs (https://archive.ismrm.org/2021/0865.html). They can be used with the --cmc and --asymmetric options respectively. Data and trained models are available here: https://zenodo.org/record/7153362

Dataset files are in raw and weights and results are in experiments. Results can be replicated using bash scripts (sac_auto_train[_cmc|_asym|_cmc_asym].sh) in the scripts folder of the code. The DATASET_FOLDER variable must be initialized to the folder where the raw and experiments folders are.

Track-to-Learn (2021)

Théberge, A., Desrosiers, C., Descoteaux, M., & Jodoin, P. M. (2021). Track-to-learn: A general framework for tractography with deep reinforcement learning. Medical Image Analysis, 72, 102093.

The reference commit to the main branch for this work is e5f2e6008e499f46af767940b5b1eec7f9293859.

See published version: https://www.sciencedirect.com/science/article/pii/S1361841521001390

See preprint: https://www.biorxiv.org/content/10.1101/2020.11.16.385229v1

A bug in the original implementation prevents the reproduction of the published results. The results presented in "What matters in Reinforcement Learning for Tractography" should be reproducible and considered as reference.

How to cite

If you want to reference this work, please use (at least) one of

@article{theberge2024matters,
  title={What matters in reinforcement learning for tractography},
  author={Th{\'e}berge, Antoine and Desrosiers, Christian and Bor{\'e}, Arnaud and Descoteaux, Maxime and Jodoin, Pierre-Marc},
  journal={Medical Image Analysis},
  volume={93},
  pages={103085},
  year={2024},
  publisher={Elsevier}
}
@article{theberge2021,
title = {Track-to-Learn: A general framework for tractography with deep reinforcement learning},
journal = {Medical Image Analysis},
pages = {102093},
year = {2021},
issn = {1361-8415},
doi = {https://doi.org/10.1016/j.media.2021.102093},
url = {https://www.sciencedirect.com/science/article/pii/S1361841521001390},
author = {Antoine Théberge and Christian Desrosiers and Maxime Descoteaux and Pierre-Marc Jodoin},
keywords = {Tractography, Deep Learning, Reinforcement Learning},
}