Installation

1. Create a conda environment (recommended)

ENVNAME="jstablenv"
conda create -n $ENVNAME python -y
conda activate $ENVNAME

2. Install PyTorch

Please install PyTorch for your CUDA toolkit within the conda environment:

3. Install jSTABL

Within the conda environment:

(jstablenv):~ pip install -e  git+https://github.com/ReubenDo/jSTABL#egg=jSTABL

Data

Control data

Consists of 35 T1 scans from the OASIS project with annotations of 143 structures of the brain provided by Neuromorphometrics, Inc. under academic subscription. From the 143 structures, we deducted the 6 tissue classes. Additionnaly, 25 T1 control scans from ADNI-2 were added with bronze standard parcellation of the brain structures computed with the accurate but time-consuming GIF algorithm. The T1c, T2 and FLAIR scans are missing.

Glioma data

Consists of 285 patients (210 with high grade glioma and 75 with low grade glioma) from the training set of BraTS18. T1, T1c, T2 and FLAIR scans are provided for each patient. Three tumour structures are annotated. The tissue annotations are missing.

White Matter Lesions data

Consists of 60 sets of brain MR imagesfrom the White Matter Hyperintensities (WMH) database. T1 and FLAIR scans are provided for each patient. The white matter lesions are annotated. The tissue annotations are missing.

Training the models

Glioma

Without Domain Adaptation:

(jstablenv):~ python3 glioma/train_WMH_noDA.py --model_dir models/WMH/noDA/ 

With Data Augmentation:

(jstablenv):~ python3 glioma/train_BRATS_augmentation.py --model_dir models/BRATS/augm/

With Adversarial Domain Adaptation:

(jstablenv):~ python3 glioma/train_BRATS_adversarial.py --model_dir models/BRATS/adv/ 

With Annotated Pseudo-Healthy Scans:

(jstablenv):~ python3 glioma/train_BRATS_pseudohealthy.py --model_dir models/BRATS/pseudohealthy/ 

White Matter Lesions

To train:

(jstablenv):~ python3 wmh/train_BRATS_noDA.py --model_dir models/BRATS/noDA/