Deep Learning-based Re-identification and Anonymization of Chest Radiographs
Chest X-rays are hugely valuable diagnostically, catching diseases and abnormalities early, and radiologists increasingly rely on computer-aided detection systems built on deep learning to help. Those systems, in turn, depend on large public chest X-ray datasets for training and reproducible research. But chest radiographs also carry sensitive, potentially patient-specific information, so they must be anonymized before public release, a process that until now has focused almost entirely on stripping sensitive metadata. That leaves untouched the biometric patterns baked into the image itself, a source of identifiable information that conventional anonymization misses entirely.
This thesis investigates just how exploitable that gap really is. It shows that deep learning can automatically extract biometric “fingerprints” from chest radiographs, enabling linkage attacks that could re-identify patients in supposedly anonymous public datasets. Using siamese neural networks trained to verify or re-identify patients through one-to-one or one-to-many image comparisons, the thesis achieves 99.4% AUC and 99.6% Precision@1, demonstrating that such attacks are genuinely feasible and that current anonymization practices fall well short.
To address this, the thesis proposes two approaches for anonymizing the images themselves. The first, PriCheXy-Net, combines a U-Net, an auxiliary classifier, and a verification network to learn targeted image deformations that hide biometric patterns while preserving diagnostic usefulness. This cuts patient verification performance from 81.8% down to 57.7% AUC, with only a small hit to abnormality classification accuracy, showing a favorable privacy-utility trade-off.
The second approach generates fully synthetic chest radiographs using a class-conditional latent diffusion model. To avoid these generative models simply memorizing and reproducing real patients’ patterns, the thesis introduces a privacy-enhancing sampling strategy that keeps patient-specific biometric details out of the generated images. The resulting synthetic radiographs retain strong diagnostic utility, with only a 3.5% AUC gap compared to real data on an abnormality classification task, and a Turing test confirms that experienced radiologists can barely tell them apart from real scans.
The thesis also released open-source code for both approaches, aimed at supporting future work in this area.