Setup

• This repository requires Python >= 3.6 with pip, CUDA >= 11 and CuDNN SDK installed. CUDA and CuDNN can be installed together via apt (sudo apt install libcudnn8).

• install all required python dependencies. Therefore, in the root directory of the repository, execute:

pip3 install -r requirements.txt

• Prepare your dataset. Optional arguments are:
• -h, --help - show brief help
• -i, --inflate - Inflate the original dataset by rotating the given MRI images

./prepare_trainingset.sh [-h|--help] [-i|--inflate]

Survival Prediction - Training and Evaluation

• To train the survival prediction with regression on the BraTs dataset, execute:

./train_surv_pred [-h] <model_name> [arguments]

• To evaluate the survival prediction with regression on the BraTs dataset, execute:

./eval_surv_pred [-h] <model_name> [arguments]

• To validate the survival prediction with regression on the BraTs dataset and generate a csv file for validation on the Penn Imaging Portal, execute:

./val_surv_pred [-h] <model_name> [arguments]

• The following model names for <model_name> are currently available:
• regression_l5 - Takes one centered layer of every MRI image type (including the ground truth from the Segmentation layer) and trains a regression onto it (output is normalized [0,1]).
• regression_3d <mri_type> - Takes the 3D MRI image of type <mri_type> as input and trains a regression onto it (output is normalized [0,1]). Valid types for <mri_type> are: flair, t1, t1ce, t2
• regression_3d_scaled <mri_type> - Takes a scaled 3D MRI image (half the size of the original) of type <mri_type> as input and trains a regression onto it (output is normalized [0,1]). Valid types for <mri_type> are: flair, t1, t1ce, t2.

• regression_3dcnn <mri_type> - Takes a scaled 3D MRI image (half the size of the original) of type <mri_type> as input and trains a regression onto it using 3D convolutional feature detectors (output is normalized [0,1]). Valid types for <mri_type> are: flair, t1, t1ce, t2.

• The following optinal arguments are implemented:

• [-h] - Prints a short help
• [-c] - Continues training on the existing .h5 model (if it exists and only for training)

Troubleshooting

• If there is an error while installing the Pyhton dependencies, try upgrading pip:

pip3 install --upgrade pip

python3 -m pip install --upgrade setuptools

• Encountered Errors:

ERROR: launchpadlib 1.10.13 requires testresources, which is not installed.

Fix: sudo apt install python3-testresources

SHAP

To create shap validation images, use the shap_test.py script and adapt the required fields for file and network selection.