Packhäuser K (2025)
Publication Language: English
Publication Type: Thesis
Publication year: 2025
URI: https://open.fau.de/handle/openfau/37252
Chest radiography has enormous diagnostic value in the clinical routine, as it enables the early detection of various diseases and abnormalities in the thoracic region. In practice, radiologists are supported by computer-aided detection (CAD) systems, which increasingly rely on automated image processing algorithms based on deep learning (DL) techniques. The development of such systems significantly benefits from large public chest X-ray datasets, promoting open and direct data access, as well as transparency and reproducibility of research in the scientific community. Chest radiographs include sensitive information that may be unique to specific patients. Consequently, it is crucial to anonymize them prior to any potential public release. This process is strictly governed by various legal regulations with the overarching objective of protecting personal data and, thus, ensuring patient privacy. The conventional anonymization profiles for chest radiographs that have been widely applied to date focus almost exclusively on removing sensitive information from accompanying metadata. However, this does not affect biometric patterns in the image domain itself, representing a remaining source of personally identifiable information. With a specific emphasis on patient privacy, this thesis investigates the capabilities of DL techniques for the verification and re-identification of patients from available chest radiographs. More specifically, it demonstrates that DL can be leveraged to automatically extract biometric fingerprints from chest radiographs, enabling potentially harmful linkage attacks on public datasets. The employed threat models are represented by siamese neural networks (SNNs), which are trained to verify or re-identify individual patients by performing either 1-to-1 or 1-to-many image comparisons. With performance scores of 99.4% in the AUC and 99.6% in the Precision@1, this work indicates the feasibility of DL-based linkage attacks and, thus, highlights the urgent need for more sophisticated chest X-ray anonymization techniques. This thesis proposes two promising approaches for anonymizing the image domain of chest radiographs. First, “PriCheXy-Net” is presented – a network architecture, which aims to conceal existing biometric patterns while preserving the utility of the data. PriCheXy-Net is composed of a U-Net, an auxiliary classifier, and a verification network, which altogether allow for learning targeted image deformations, effectively impeding the success of DL-based linkage attacks. Experimental results show that the learned data modifications reduce the patient verification performance (AUC) from 81.8% to 57.7% with minimal effect on an abnormality classification task, indicating the positive impact in optimizing the so-called privacy-utility trade-off. The second anonymization approach utilizes a class-conditional latent diffusion model (LDM) to produce high-quality synthetic chest radiographs. To particularly avoid the memorization effect of generative models, this thesis proposes a privacy-enhancing image sampling strategy, ensuring that patient-specific biometric patterns are not reproduced during image generation. The resulting anonymous chest radiographs exhibit a high degree of data utility, which is verified by a small performance gap (AUC) of only 3.5% on an abnormality classification task compared to real data. The image quality is further assessed with a Turing test, demonstrating that experienced radiologists can barely distinguish the LDM-based images from real scans. Lastly, as part of this thesis, open-source code repositories were created to stimulate potential future projects and strengthen ongoing research in the explored field.
APA:
Packhäuser, K. (2025). Deep Learning-based Re-identification and Anonymization of Chest Radiographs (Dissertation).
MLA:
Packhäuser, Kai. Deep Learning-based Re-identification and Anonymization of Chest Radiographs. Dissertation, 2025.
BibTeX: Download