Privacy Risks of Whole-Slide Image Sharing in Digital Pathology

Abstract

Access to large volumes of so called whole-slide images, high-resolution scans of complete pathological slides, has become a cornerstone of development of novel artificial intelligence methods in digital pathology, but has broader impact for medical research and education/training. However, a methodology based on risk analysis for sharing such imaging data and applying the principle “as open as possible and as closed as necessary” is still lacking. In this article we develop a model for privacy risk analysis for whole-slide images, which focuses primarily on identity disclosure attacks, as these are the most important from a regulatory perspective. We develop a mathematical model for risk assessment and design a taxonomy of whole-slide images with respect to privacy risks. Based on these risk assessment model and the taxonomy, we design a series of experiments to demonstrate the risks on real-world imaging data. Finally, we develop guidelines for risk assessment and recommendations for data sharing based on identified risks to promote low-risk sharing of the data.

Digital Pathology, Deep Learning, and Cancer: A Narrative Review

“…As WSI deals with large amounts of sensitive patient information, strong data privacy policies, encryption, and access control are indispensable. Differential privacy methods such as encryption and federated learning protect individual privacy while allowing accurate data analysis (87–89).…”

Anonymization of whole slide images in histopathology for research and education

“…Preferably, the content of these image tiles should be arbitrary, should not overlap, and should not have prominent edges, to reduce mutual information between tiles. Even though this already significantly reduces the traceability, this approach does not ensure entire anonymity, as image matchers might still be able to re-identify the respective WSI as described by Holub et al 21 However, the main limitation of Holub's work is that only (a) the entire downsampled WSI and (b) two equally sized regions of variably overlapping tissue regions are considered for the similarity comparison. The study does not address the case of locating an arbitrary image region in the entire search space of a WSI or even a large dataset of WSIs.…”

“…Thus, an individual patient can be identified by appropriate comparison between the WSI and the original slides. Holub et al 21 have published a preprint describing an attack scenario in which the attacker can access an identifying WSI dataset based on background information through feature extraction and similarity comparison of image details. The scenario shown outlines a linkage attack on similar and consecutive WSIs and focuses on assessing the likelihood of such an attack.…”