A novel approach of adding bone information to the 4-class PET/MR attenuation correction method commonly used in clinical practice. The metod directly predicts the error in the PET image of bones not being included in the 4-class attenuation correction maps. For more information see please read our paper: Pelvic PET/MR attenuation correction in the image space using deep learning.

The easiest way to use the code is through the supplied dockerfile. See the supplied docker README.md for instructions. If visual studio code is used, the supplied devcontainer can be used to automatically set up a development environment inside the docker container.

The preprocessing of the data is performed as described in the paper. The data is expected to be stored in the following folder structure

datasetName
├── train
│   ├── A
│   └── B
└── val
    ├── A
    └── B

The images in A and B are both saved numpy arrays where each saved file contains one image slice. Normally each file in A would be a (256,256,3) matrix and B a (256,256) matrix. A working dummy dataset can be created by running scripts/create_debug_dataset.py from the project root folder.

The commands used for training and testing the best performing model can be found in the script scripts/resnet_9blocks.sh. The training is monitored using tensorboard.

The code is based on this excellent pytorch pix2pix implementation which again was inspired by DCGAN.