Originally designed after this paper on volumetric segmentation with a 3D U-Net. The code was written to be trained using the BRATS data set for brain tumors, but it can be easily modified to be used in other 3D applications.

All required installations can be made using the provided Dockerfile as:

docker build . -t 3dunetcnn:latest

Thereafter, clone the project and cd into the project directory:

cd 3DunetCNN

The contructed docker image is entered using:

NV_GPU=0 nvidia-docker run --security-opt seccomp=unconfined --rm -u $(id -u):$(id -g) -v $(pwd):$(pwd) -w $(pwd) -ti 3dunetcnn:latest /bin/bash

where NV_GPU is used to select the GPU to train with. After entering the /bin/bash you need to add the working folder to pythonpath by:

export PYTHONPATH=${PWD}:$PYTHONPATH

Train the network using either train.py for the original 3D U-net and train_isensee2017.py for the improved version. These files are in the different folders for different data scenarios: brats, headneck and iseg. Instructions for downloading and preprocessing the data are provided in the corresponding folders. One example of initiating training on the brats data using the original 3D U-net can be seen below:

python brats/train.py

Without using docker, follow the tutorial for the Brats data for installation and testing

See README in the brats folder