Privacy Risks in the Anonymization of Medical Image Data

Publicly available medical image datasets are essential for the progress of supporting diagnostic algorithms in radiology. In terms of patient data protection, it is necessary to anonymize images before publication. However, there is a risk that these anonymization procedures are insufficient. In this study, we employ a specially developed siamese neural network to assess the ability of re-identifying additional X-ray images of specific patients within supposedly anonymized datasets. This is analyzed using different neural network constellations and images from two variable datasets, CheXpert and KI-Rad-MSK, which include chest and wrist radiographs. Our results show that conventional anonymization approaches cannot withstand attacks using modern deep learning methods: One image of a patient is sufficient to re-identify other images of the same person within large datasets. Instead, we recommend innovative variants, such as latent diffusion models, to ensure data protection without compromising progress in medical imaging.