PriCheXy-Net

Overview

This repository provides code to the paper "Deep Learning-Based Anonymization of Chest Radiographs: A Utility-Preserving Measure for Patient Privacy" (accepted at MICCAI 2023).

In this work, we proposed PriCheXy-Net — an adversarial image anonymization architecture for chest radiographs (see below). The model is a composition of three independent neural networks that, when collectively used, allow for learning a deformation field that is able to impede patient re-identification.

Requirements

All our experiments have been conducted using:

• Python (3.9.5)
• PyTorch (1.10.2)

A list of required packages can be found in requirements.txt.

Usage instructions

1. Pre-training of the flow field generator

Set the experiment configuration in ./config_files/config_pretrain.json. The most important parameters are:

• "generator_type": Either "flow_field" (ours) or "privacy_net".
• "mu": The deformation degree, e.g. 0.01.

Execute the pre-training by running the following command:

python3 pretrain_generator.py --config_path ./config_files/ --config config_pretrain.json

Note that we already provide the pre-trained models in folder ./networks/.

2. Training of PriCheXy-Net

Set the experiment configuration in ./config_files/config_anonymization.json. The most important parameters are:

• "ac_loss_weight": Set to 1 for all our experiments.
• "ver_loss_weight": Set to 1 for all our experiments.
• "generator_type": Either "flow_field" (ours) or "privacy_net".
• "mu": The deformation degree, e.g. 0.01.

Execute the complete training by running the following command:

python3 train_architecture.py --config_path ./config_files/ --config config_anonymization.json

Note that we provide the trained models in folder ./networks/.

3. Re-training and evaluation of the verification model

Set the experiment configuration in ./config_files/config_retrainSNN.json. The most important parameters are:

• "perturbation_type": Either "flow_field" (ours), "privacy_net", "dp_pix" or "none".
• "perturbation_model_file": The perturbation model, e.g. "./networks/generator_lowest_total_loss_mu_0.01.pth".
• "mu": The deformation degree, e.g. 0.01.
• "b": The pixelization grid size, e.g. 8. Only needed for "dp_pix".
• "m": Set to 1 for our experiments. Only needed for "dp_pix".
• "eps": Set to 0.1 for our experiments. Only needed for "dp_pix".

To assess the anonymization capability, run the following command:

python3 retrain_SNN.py --config_path ./config_files/ --config config_retrainSNN.json

Note that this script only performs one training and testing run of the SNN! In our paper, however, we report the verification performance (AUC (mean +- std)) over 10 independent training/testing runs. If this experiment is conducted multiple times, one can compute the average verification performance using the function compute_meanAUC() in ./utils/utils.py.

4. Evaluation of the classification model on anonymized data

Set the experiment configuration in ./config_files/config_eval_classifier.json. The most important parameters are:

• "perturbation_type": Either "flow_field" (ours), "privacy_net", "dp_pix" or "none".
• "perturbation_model_file": The perturbation model, e.g. "./networks/generator_lowest_total_loss_mu_0.01.pth".
• "mu": The deformation degree, e.g. 0.01.
• "b": The pixelization grid size, e.g. 8. Only needed for "dp_pix".
• "m": Set to 1 for our experiments. Only needed for "dp_pix".
• "eps": Set to 0.1 for our experiments. Only needed for "dp_pix".

To assess data utility on a downstream classification task, run the following command:

python3 eval_classifier.py --config_path ./config_files/ --config config_eval_classifier.json

To compute the 95% confidence intervals (1000 bootstrap runs), use the function bootstrap_abnormalities() in ./utils/utils.py.