similarity

Linkage attacks on whole-slide images (WSIs): given a slide, find the slides of the same patient by comparing image features. The code was written by RationAI (Masaryk University) and partners for PRIVAGAMS, a project funded in the 1st OSCARS Open Call (grant agreement No. 101129751), and for the public challenge Benchmarking WSI Privacy.

What is where

Path	Content
src/	Slide-similarity pipeline: tile embeddings (pathology foundation models) aggregated into slide-level descriptors (mean pooling, Fisher vectors, VLAD) for retrieval and re-identification experiments.
notebooks/analyze.ipynb	Embedding-quality analysis: re-identification rates, cross-scanner robustness, preprocessing and mpp impact, scaling.
memorisation/	Memorisation audit of synthetic patches against training tiles (UNI2-h features, FAISS), and an edge-based near-duplicate check. See its README.
steganography/	Script that hides one image inside another. See its README.

Pipeline

1. Tile embeddings — src/tile_embeddings.py Tiles each WSI at a given mpp, optionally removes background, normalizes H&E staining and/or applies CLAHE, then embeds 224×224 tiles with a pathology foundation model. Writes one parquet shard set per slide (slide_id=<encoded path>/ partitions).

2. Slide embeddings — src/slide_embeddings.py Aggregates tile embeddings into fixed-length slide descriptors: mean pooling, Fisher vectors (GMM) and VLAD (K-Means), each computed on raw and spatially-smoothed "super tile" embeddings, over multiple seeded runs.

3. Analysis — notebooks/analyze.ipynb Embedding-quality experiments: re-identification success rates, cross- scanner robustness, preprocessing and mpp impact, and scaling analysis.

Encoders

--encoder	Model
0	GigaPath
1	Virchow2
2	UNI2-h
3	midnight-12k
4	SimCLRv2 (ResNet152x4)

Weights are downloaded from the Hugging Face Hub on first use.

Installation

Python 3.12 or newer. The dependencies are in pyproject.toml and locked in uv.lock:

uv sync

OpenSlide and libvips must be installed on the system. The memorisation/ scripts are standalone and document their own dependencies (FAISS, fire, …); they are not part of the similarity package.

Setup (Musica cluster)

mamba create -n similarity-env python=3.12.9 -y
mamba activate similarity-env
mamba install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidia -y
mamba install -c conda-forge libvips openslide libgcc-ng -y
uv pip install -e .

Usage

All commands are run from the repository root.

# 1. tile embeddings for a folder of WSIs (encoder 1 = Virchow2)
python -m src.tile_embeddings --slide-path /data/slides --save-path /data/tile_embeddings --encoder 1

# 2. slide descriptors (mean/Fisher/VLAD, several seeded runs)
python -m src.slide_embeddings --slide-path /data/tile_embeddings --save-path /data/slide_embeddings --num-runs 20

Preprocessing toggles for tile embeddings: --rmbg (drop non-tissue tiles), --normalize (stain normalization), --clahe (contrast enhancement), --mpp (tiling resolution, default 0.5 = 20×), --overwrite.

Related resources

• Challenge: https://www.kaggle.com/competitions/benchmarking-wsi-privacy
• Challenge data: BBMRI-ERIC WSI archive, https://cbioportal.wsi-archive.bbmri-eric.eu/archive/ (Life Science Login required).
• Risk assessment this code builds on: J. Buchta, Assessing Privacy Risks in Real and Synthetic Whole-Slide Images via Foundation Models, diploma thesis, Masaryk University, 2026, which extends Holub et al., Privacy risks of whole-slide image sharing in digital pathology, Nature Communications 14, 2577 (2023), https://doi.org/10.1038/s41467-023-37991-y.

Licence

This project is licensed under the MIT License.