This repository contains the code from the paper "On instabilities of deep learning in image reconstruction - Does AI come at a cost?", by V. Antun, F. Renna, C. Poon, B. Adcock and A. Hansen.

In order to make this code run you will have to download and install the neural networks we have considered. Most of the necessary data can be downloaded from https://www.mn.uio.no/math/english/people/aca/vegarant/data/storage2.zip (4.9 GB). Please note that you will have to modify all paths in the source files so that they point to the data. You will also need to add the directory py_adv_tools to your python path.

For the state of the art reconstruction we have used the ShearletReweighting code from J. Ma & M. März paper and spgl1. These repositories must also be downloaded and added to your Matlab path.

To test the FBPConvNet you will have to download and install MatConvNet and the FBPConvNet and add these repositories on you matlab path. From within the invfool/FBPConvNet directory you should then be able to run the scripts.

Download and install the DeepMRINet and add it to your pythonpath. Note that to run DeepMRINet you need a very specific version of Theano and Lasagne. See the GitHub page of DeepMRINet for more information about this. Then run the code in the DeepMRI folder.

Download the network code for MRI-VN and add it to your python path. Note that this network requires a custom-made version of tensorflow, tensorflow-icg. To run "the add more samples" experiment you need to download the data from GLOBUS.

These experiments are self contained. It requires a vanilla Tensorflow install.

The original code for this network can be found at this GitHub page. Full dataset can be downloaded from MICCAI 2013 grand challenge page. We provide all code and data necessary to reproduce the figures in the paper. We do not provide the code to train the network nor the full dataset, as this can be found via the links above. To run the code you need Tensorflow and Tensorlayer.