Reconstructing magnetic resonance (MR) images from undersampled k-space measurements can potentially decrease MR examination times. This repository contains thesource code for a hybrid k-space/image domain model proposed by our group. Our model is called W-net. If you use this code in your experiments, we ask you to kindly cite our paper:
The MR raw-data is publicly available as part of the Calgary-Campinas dataset for benhcmarking purposes.
The code was developed using Python 2.7, NumPy, TensorFlow and Keras. It should be easy to use it, but we appreciate any feedback on how to improve our repository.
W-net architecture. It is composed of a residual U-net on k-space domain connected to an image domain U-net through the magnitude of the inverse discrete fourier transform operator.
Sample reconstruction of our W-net and four other methods published in the literature with special highlight on the cerebellum region. Speed-up factor of 5x.
W-net reconstruction gif. From left to right: fully sampled reconstruction, W-net reconstruction from a k-space undersample by 80%, and absolute error differences.
Any questions? roberto.medeirosdeso@ucalgary.ca
MIT License Copyright (c) 2017 Roberto M Souza