Benchmarking WSI Privacy is a challenge focused on linkage attacks and the anonymization of histopathology slides using synthetic data. Whole-slide images (WSIs) are gigapixel-scale digital scans that serve as the engine for modern AI in pathology. As these images are increasingly shared across research institutions, a critical privacy question arises: can a patient be identified from their tissue image alone, even without clinical data? Recent research suggests that microscopic tissue structure acts like a biological fingerprint, allowing computational models to link images from the same patient across different datasets.
This competition invites machine learning engineers, medical imaging researchers, privacy scientists, and curious individuals to explore this risk empirically. No pathology background is required.
Track 1 — Patient-Level Slide Linkage The goal is to identify which slides in a gallery belong to the same patient as a given probe WSI, using image content alone with no metadata. Each probe comes with a gallery of 100 candidate slides. Zero, one, or several of them may genuinely come from the probe's patient. Participants must return the subset of gallery slides they believe share the probe's patient. This track benchmarks how linkable real WSIs are across different tissue sections and scanning conditions.
Track 2 — Synthetic Anonymization The goal is to retrieve the real slide from which a synthetically generated slide was derived. A diffusion model has been used to generate synthetic WSIs conditioned on real tissue images. If a successful match is found, it indicates that the synthetic image still carries patient-identifying information, suggesting that diffusion-based generation alone may not be sufficient for true anonymization.
The dataset covers 250 patients, split into 150 for training and 100 for testing, with no overlap between splits. At launch, participants will receive a public manifest mapping anonymized slides to pseudonymous IDs and probe/gallery pairs for both tracks. The competition runs from August to October 2026, with results announced in November 2026.
Submissions are scored using mean per-example F1 between predicted matching sets and the true set. This challenge is academically motivated; top-performing teams from each track will be invited to co-author a joint publication summarizing the findings and implications for WSI data sharing policy. Certificates of achievement will also be awarded.