Brain-Tumor-Segmentation

Usage

1. Install Dependencies

pipenv install
pipenv shell

2. Set up .env

cp .env.example .env

3. Run Exp. with Command Line

python main.py -m <model_id> -d <data_provider_id> [--comet]

• Please refer to models/__init__.py for available model_ids, and data/data_providers.py for available data_provider_ids.
• For other arguments, please refer to parser.py
• Passing the --comet argument allows the user to log results to comet.ml, you'll have to add your api-key to the .env file

4. Resume Training from Checkpoint

python main.py --checkpoint_dir <checkpoint_dir>

5. Prediction

python predict.py --checkpoint_dir <checkpoint_dir> [--predict_mode] [--save_volume]

Data Configuration

NTU

0. (Recommended) If you wish to preprocess with masks, run:

bash ./preprocess_tools/skull_strip.sh <bse_dir> <result_dir> <data_dir>

Where bse_dir is the binary file from the BrainSuite package.

1. To preprocess, run:

cd preprocess_tools
python preprocess.py <data_dir> <mask_dir> <result_dir> 

• This script will resize to $1mm^3$ resolution and center the brain with its mask.
• The resulting shape is defined in the ImageProcessor class in preprocess_tools/image_utils.py

2. To postprocess, run:

python postprocess.py <lable_dir> <result_dir> <image_processor_path>

• This script will revert the resolution and copy the according headers from the original nii file.

NTU2019

0. Directory Structure:

NTU2019
    image
    label
NTU2019_result
    image
    mask

BRATS2015

0. Directory Structure:

BRATS2015
    BRATS2015_Training
    BRATS2015_Testing

• The inner structure is same as the original format downloaded from https://www.smir.ch/BRATS/Start2015.

TCIA_CT

• Data from https://github.com/deepmind/tcia-ct-scan-dataset/tree/master/nrrds
• The corrensponding options for data_provider_id are tciact_oncologist, tciact_radiographer, tciact for both

Dependencies

• BrainSuite (optional)
• python 3.6
• pipenv
• python packages listed in Pipfile are managed by pipenv