Volumetric segmentation using a V-Net model

We recommend on using the following docker image. Pull the following docker image using:

docker pull geromev/dsgzneuro:latest 

Alternatively, you can also:

pip install -r requirements.txt 

All volumes should be in ./vols_all/.

We use Segmentation3DTrainer object in trainer.py which performs data loading, data augmentation, and training. As an example usage to train a model you can do the following:

python trainer_example_usage.py --train_model=1

Training will ouptut the following:

1. Save weights of the trained model at ./model/.
2. Output tensorboard event logs at ./logs/.
3. If debug_dump_image is set to True, will save augmented volumes at ./augmented_vols.

When training is finished, the trained model will be used to evaluate the test set.. You can evaluate the test set using the following:

python trainer_example_usage.py --train_model=0

Predicted segmentation outputs will be in ./pred_output/.

This project is licensed under the MIT License - see the LICENSE.md file for details