Related Experiment Video
Updated: Sep 22, 2025
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
9.0K
Toward Sharing Brain Images: Differentially Private TOF-MRA Images With Segmentation Labels Using Generative
Tabea Kossen1,2, Manuel A Hirzel1, Vince I Madai1,3,4
1CLAIM-Charité Lab for AI in Medicine, Charité Universitätsmedizin Berlin, Berlin, Germany.
Frontiers in Artificial Intelligence
|May 19, 2022
Summary
Synthetic data generated using Generative Adversarial Networks (GANs) with differential privacy can preserve patient privacy for medical imaging tasks like brain vessel segmentation. Stricter privacy levels slightly reduce segmentation accuracy but maintain usability.
Area of Science:
- Medical Imaging
- Deep Learning
- Data Privacy
Background:
- Sharing labeled medical data is essential for Deep Learning but hindered by privacy regulations.
- Standard anonymization techniques for medical data can be reversible.
- Synthetic data offers a potential solution for privacy-preserving data sharing.
Purpose of the Study:
- To implement a Wasserstein Generative Adversarial Network (WGAN) with differential privacy for generating privacy-preserving labeled Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) image patches.
- To evaluate the utility of these synthetic images for brain vessel segmentation using a U-Net model.
- To assess the impact of varying privacy levels (epsilon, ϵ) on data utility and image similarity.
Main Methods:
- Implemented a Wasserstein GAN (WGAN) with and without differential privacy.
- Generated synthetic labeled TOF-MRA image patches.
- Trained a U-Net model on synthesized data for brain vessel segmentation.
- Evaluated segmentation performance on real patient data and used Fréchet Inception Distance (FID) to measure image similarity.
Main Results:
- Segmentation performance and image similarity (FID) decreased with stricter privacy guarantees (lower ϵ).
- The best model trained on private synthetic data achieved a Dice Similarity Coefficient (DSC) of 0.75 (ϵ = 7.4) and 0.69 on a second dataset.
- Performance became unstable and unusable below ϵ < 5 (DSC < 0.61).
Conclusions:
- Synthesized labeled TOF-MRA images with differential privacy retain predictive properties for brain vessel segmentation.
- Privacy-preserving synthetic data is a promising approach for data sharing in medical imaging.
- Further research is needed for generalizability and performance enhancement.
