OpenSRH: optimizing brain tumor surgery using intraoperative stimulated Raman histology

Companion code for OpenSRH. Paper submitted to NeurIPS 2022 Datasets and Benchmarks Track.

Paper Website / arXiv / MLiNS Lab

Installation

1. Clone OpenSRH github repo

   git clone [email protected]:MLNeurosurg/opensrh.git

2. Install miniconda: follow instructions here
3. Create conda environment

   conda create -n opensrh python=3.9

4. Activate conda environment

   conda activate opensrh

5. Install package and dependencies

   <cd /path/to/opensrh/repo/dir>
   pip install -e .

Directory organization

• opensrh: the library for training with OpenSRH
  • datasets: PyTorch datasets to work with the data release
  • losses: loss functions for contrastive learning
  • models: PyTorch networks for training and evaluation
  • train: training and evaluation scrpits
• README.md
• LICENSE

Training / evaluation instructions

The code base is written using PyTorch Lightning, with custom network and datasets.

Cross entropy experiments

1. Download and uncompress the data.
2. Update the sample config file in train/config/train_ce.yaml with desired configurations.
3. Change directory to train and activate the conda virtual environment.
4. Use train/train_ce.py to start training:

   python train_ce.py -c config/train_ce.yaml

Contrastive learning experiments

1. Download and uncompress the data.
2. Update the sample config file in train/config/train_contrastive.yaml with desired configurations.
3. Change directory to train and activate the conda virtual environment.
4. Use train/train_contrastive.py to start training:

   python train_contrastive.py -c config/train_contrastive.yaml

5. To run linear or finetuning protocol, update the config file train/config/train_finetune.yaml and continue training using train/train_finetune.py:

   python train_finetune.py -c config/train_finetune.yaml

Model evaluation

1. Update the sample config file in train/config/eval.yaml with desired configurations, including the PyTorch Lightning checkpoint you would like to use.
2. Change directory to train and activate the conda virtual environment.
3. Use train/train_ce.py to start training:

   python eval.py -c config/eval.yaml

License Information

OpenSRH data is released under Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0), and the code is licensed under the MIT License. See LICENSE for license information and THIRD_PARTY for third party notices.