Description

This repository contains the code to reproduce our experiments for the paper Triggering hallucinations in model-based MRI reconstruction via adversarial perturbations (to appear).

Setup

Install the required dependencies using pip:

pip install -r requirements.txt

Data set

Our experiments make use of the fastMRI data set, which is publicly available. We only evaluated our attack on the brain and knee portions of the data, but prostate and breast data should work just as well.

CLI reference

There are three scripts in the codebase meant for external use:

• main.py, which computes all the adversarial perturbations
• report.py, which computes all the metrics
• display.py, which plots the metrics
• visualize.py, which visualizes the resulting MR images

Once the environment has been set up, experiments can be run using the main.py script. This script takes a number of arguments, detailed below:

Argument	Type	Default	Description
data	string	—	Path to the fastMRI data set.
--out PATH	string	out	Path to the output directory where the generated perturbations shall be stored.
--model {unet,varnet}	string	unet	Model to target (either UNet or E2E-VarNet).
--organ {knee,brain}	string	knee	Organ data to use (either knee or brain).
--coil {sc,mc}	string	sc	Use single-coil (sc) or multi-coil (mc) data.
--shape {line,square}	string	line	Shape to draw on the target images.
--iterations INT	integer	150	Maximum number of iterations in the attack.
--eps FLOAT	float	1e-6	Perturbation budget of the attack.
--step FLOAT	float	1e-7	Attack step size.

Note that our attack can also use squares as the target shapes, although these results were not reported in the paper since we find them to be inferior to lines.

The main.py script generates the perturbations and stores them in the specified output directory. Once this script is completed, a report can be generated using report.py, which takes the same arguments as main.py. These reports take the form of CSV files containing all measurements of the metrics detailed in the paper.

To generate the plots we showed in the paper, run the display.py script after generating the report.