gschramm / pyapetnet

a CNN for anatomy-guided deconvolution and denoising of PET images
https://gschramm.github.io/pyapetnet/
MIT License
16 stars 5 forks source link
anaconda cnn miniconda

pyapetnet

A convolutional neurol network (CNN) to mimick the behavior of anatomy-guided PET reconstruction in image space.

architecture of pyapetnet

Authors

Georg Schramm, David Rigie

License

This project is licensed under the MIT License - see the LICENSE file for details.

Scientific Publication

Details about pyapetnet are published in Schramm et al., "Approximating anatomically-guided PET reconstruction in image space using a convolutional neural network" ,NeuroImage Vol 224 2021. If we you are using pyapetnet in scientific publications, we appreciate citation of this article.

Installation

We recommend to install pyapetnet from conda-forge via

conda -c conda-forge install pyapetnet

Alternatively, pyapetnet can be also installed from pypi via

pip install pyapetnet

Testing the installation

To test the installation activate your virtual environment and run

import pyapetnet
print(pyapetnet.__file__)

If the installation was successful, a number of command line scripts all starting with pyapetnet* to e.g. do predictions with the included trained models from nifti and dicom input images will be available.

Getting started - running your first prediction with pre-trained models

To run a prediction using one of included pre-trained networks and nifti images, run e.g.:

pyapetnet_predict_from_nifti osem.nii t1.nii S2_osem_b10_fdg_pe2i --show

Use the following to get information on the (optional) input arguments

pyapetnet_predict_from_nifti -h

To get a list of available pre-trained models run

pyapetnet_list_models

To make predictions from dicom images, use

pyapetnet_predict_from_dicom osem_dcm_dir t1_dcm_dir S2_osem_b10_fdg_pe2i --show

The source code of the prediction scripts can be found in the scripts subfolder here.

Training your own model

If you want to train your own model (from scratch or using transfer learning) using your own data, have a look at our training script. All input parameters (e.g. data sets to use) have to be specified in a config json file (example here). The input data sets have to be in nifti format and should be aligned already.