[Paper Review] A Deep Learning Approach to Private Data Sharing of Medical Images Using Conditional GANs

Hanxi Sun, Jason Plawinski+8 more|arXiv (Cornell University)|Jun 24, 2021

AI in cancer detection35 references4 citations

TL;DR

This paper proposes a conditional GAN-based method to generate synthetic MRI images of vertebral units while preserving privacy and clinical relevance. By conditioning the generator on anatomical location (cervical, thoracic, lumbar), the approach produces high-fidelity, diverse, and privacy-preserving synthetic data that supports downstream tasks like model training with performance comparable to real data.

ABSTRACT

Sharing data from clinical studies can facilitate innovative data-driven research and ultimately lead to better public health. However, sharing biomedical data can put sensitive personal information at risk. This is usually solved by anonymization, which is a slow and expensive process. An alternative to anonymization is sharing a synthetic dataset that bears a behaviour similar to the real data but preserves privacy. As part of the collaboration between Novartis and the Oxford Big Data Institute, we generate a synthetic dataset based on COSENTYX (secukinumab) Ankylosing Spondylitis clinical study. We apply an Auxiliary Classifier GAN to generate synthetic MRIs of vertebral units. The images are conditioned on the VU location (cervical, thoracic and lumbar). In this paper, we present a method for generating a synthetic dataset and conduct an in-depth analysis on its properties along three key metrics: image fidelity, sample diversity and dataset privacy.

Motivation & Objective

• To address the challenge of sharing sensitive medical imaging data due to privacy concerns in clinical research.
• To develop a synthetic data generation method that preserves clinical relevance while ensuring patient privacy.
• To evaluate synthetic data on three core metrics: image fidelity, sample diversity, and dataset privacy.
• To demonstrate that synthetic data can effectively substitute real data for training and validating deep learning models in medical imaging.
• To provide a reproducible, open-source framework for generating and sharing synthetic medical images.

Proposed method

• An Auxiliary Classifier GAN (AC-GAN) is trained on real MRI images of vertebral units, conditioned on anatomical location (cervical, thoracic, lumbar).
• The generator takes a noise vector z ∈ ℝ¹⁰⁰ and a one-hot encoded condition vector cₖ ∈ ℝ³ to generate synthetic images.
• Image morphing is achieved via linear interpolation between condition vectors, enabling smooth transitions between anatomical regions.
• Privacy is evaluated using pairwise and distribution-based re-identification attacks in pixel space and embedding space (via UMAP-64).
• A synthetic dataset of 10,000 images is generated with class distribution matching the real data (25% cervical, 55% thoracic, 20% lumbar).
• A ResNet-18 classifier is trained on both real (F2305) and synthetic data, with performance compared on an independent test set (A2209).

Experimental results

Research questions

• RQ1 Can a conditional GAN generate synthetic MRI images of vertebral units that are visually and statistically similar to real images?
• RQ2 Does the synthetic dataset preserve the diversity of anatomical variations seen in the original data?
• RQ3 To what extent can the synthetic data prevent re-identification of original patients through pairwise or distribution-based attacks?
• RQ4 Can a deep learning model trained on synthetic data achieve comparable performance to one trained on real data?
• RQ5 Is the synthetic dataset a viable substitute for real data in downstream clinical image analysis tasks?

Key findings

• The synthetic images exhibit high visual fidelity, with no significant differences in perceptual quality compared to real images as assessed by human and model-based evaluations.
• The synthetic dataset captures the full range of anatomical variation across cervical, thoracic, and lumbar regions, demonstrating strong sample diversity.
• No original image was successfully re-identified in pairwise attacks, with minimum distances in both pixel and embedding spaces exceeding re-identification thresholds.
• The distribution-based attack results showed that no synthetic sample was within the 1st percentile (pixel space) or 0.1st percentile (embedding space) of any real sample, indicating strong privacy preservation.
• The ResNet-18 classifier trained on synthetic data achieved an AUC of 0.89 on the independent test set, closely matching the real-data-trained model's AUC of 0.90.
• The code and synthetic dataset are publicly available at https://github.com/tcoroller/pGAN, enabling reproducibility and further research.