Abstract
While Coronary Angiography is essential for AI diagnostics, ensuring data privacy within healthcare Information Systems remains challenging, as conventional anonymization fails to eliminate hardware-specific forensic signatures. Addressing the problem of how to balance strict privacy with downstream diagnostic utility, we propose a novel obfuscation pipeline. Instead of traditional image editing, we synthesize angiograms derived solely from binary vessel masks. This removes forensic artifacts while retaining the vascular structures vital for clinical diagnosis. We evaluate this against an Attribute Inference Attack identifying institutional data sources, comparing Pix2Pix and Diffusion architectures against classic augmentations. The proposed methods drop attack success from $F1=0.8$ to $F1\approx 0.3$ while maintaining binary segmentation $F1\approx 0.64$ (Pix2pix) and $\approx 0.79$ (Diffusion). Results demonstrate that mask-conditioned reconstruction significantly reduces the risk of source attribution, offering a means of anonymized medical data sharing across inter-organisational IS without a significant diagnostic utility trade-off.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.78
Mask-Conditioned Generative Reconstruction of Vascular Angiography as Forensic Obfuscation
While Coronary Angiography is essential for AI diagnostics, ensuring data privacy within healthcare Information Systems remains challenging, as conventional anonymization fails to eliminate hardware-specific forensic signatures. Addressing the problem of how to balance strict privacy with downstream diagnostic utility, we propose a novel obfuscation pipeline. Instead of traditional image editing, we synthesize angiograms derived solely from binary vessel masks. This removes forensic artifacts while retaining the vascular structures vital for clinical diagnosis. We evaluate this against an Attribute Inference Attack identifying institutional data sources, comparing Pix2Pix and Diffusion architectures against classic augmentations. The proposed methods drop attack success from $F1=0.8$ to $F1\approx 0.3$ while maintaining binary segmentation $F1\approx 0.64$ (Pix2pix) and $\approx 0.79$ (Diffusion). Results demonstrate that mask-conditioned reconstruction significantly reduces the risk of source attribution, offering a means of anonymized medical data sharing across inter-organisational IS without a significant diagnostic utility trade-off.
Recommended Citation
Malinowski, H., Lau, D., Szyjut, J. & Dziubich, T.(2026). Mask-Conditioned Generative Reconstruction of Vascular Angiography as Forensic Obfuscation. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.78