EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders

EXPOSE is a framework for learning explainable and domain-robust representations from pathology vision foundation model (VFM) embeddings using sparse autoencoders (SAEs).

Accepted at the eXCV Workshop of ECCV 2026.

---

Overview

EXPOSE consists of three main stages:

1. Sparse Autoencoder (SAE) training: Learn sparse features from pathology VFM embeddings.

2. Domain-specific feature identification: Train a domain classifier and identify SAE features associated with domain-specific information.

3. Domain-robust classifier training: Train the downstream relapse classifier while masking the top-$k$ domain-specific SAE features.

The resulting representations aim to preserve task-relevant information while reducing sensitivity to domain-specific variation.

---

Environment

Create the conda environment:

conda create --name expose python=3.11
conda activate expose

Install the required Python packages:

pip install -r requirements.txt

---

Data

EXPOSE operates on embeddings extracted from a pathology vision foundation model (VFM).

The code requires:

1. Precomputed VFM embeddings
2. A metadata CSV containing domain and downstream task information

VFM Embeddings

We use H0-mini (Hugging Face Link) to extract embeddings from TMA spot images.

The embeddings should have a shape of:

n_patches × embedding_dim

and each image patch features should be stored as a separate HDF5 file in data/embeddings/ like this:

import h5py
embedding_path = "data/embeddings/..."
with h5py.File(embedding_path, 'w') as file:
   file.create_dataset("patch_features", data = patch_features)

For example:

data/
├── embeddings/
│   ├── sample_001.h5
│   ├── sample_002.h5
│   └── ...
└── metadata.csv

Each embedding file should contain the VFM embeddings corresponding to the respective sample/image.

Metadata

A metadata CSV is required at:

data/metadata.csv

The metadata must contain information for each sample/image, including at least the following columns:

Column	Description
sample_id	Unique sample/image identifier
domain	Domain, site, or cohort identifier
domain_label	Domain/site/cohort label
relapse_label	Downstream relapse task label
embedding_path	Path to the corresponding embedding file
split	Dataset split (e.g., train, val, or test)

The identifiers and paths specified in metadata.csv must correspond to the available embedding files. In particular, every embedding_path must point to an existing embedding, and each embedding should be associated with the appropriate sample_id.

Example

A metadata file may look like:

sample_id,domain,domain_label,relapse_label,embedding_path,split
0,domain1,0,0,sample_0.h5,train
1,domain1,0,0,sample_0.h5,val
2,domain2,1,1,sample_1.h5,train
3,domain2,1,1,sample_1.h5,val

---

Directory Structure

After setup, the repository should have a structure similar to:

EXPOSE/
├── data/
│   ├── embeddings/
│   │   ├── ...
│   └── metadata.csv
│
├── scripts/
│   ├── 01_run_train_sae.sh
│   ├── 02_run_train_domain_classifier.sh
│   ├── 03_run_define_domain_features.sh
│   └── 04_run_train_relapse_classifier.sh
│
├── src/
│
├── requirements.txt
└── ... (other files)

---

EXPOSE Training

1. Sparse Autoencoder Training

First, train a sparse autoencoder (SAE) on the VFM embeddings.

Run:

bash scripts/01_run_train_sae.sh

The SAE learns a sparse representation of the original VFM embeddings.

The resulting SAE features are used in the subsequent domain analysis and downstream classification steps.

---

2. Domain-Specific Feature Identification

The second stage identifies SAE features that encode domain-specific information.

2.1 Train the Domain Classifier

Train a classifier to predict the domain from the SAE representation:

bash scripts/02_run_train_domain_classifier.sh

The classifier is used to determine which SAE features contain information predictive of the domain.

2.2 Define Domain-Specific Features

Obtain linear weights from domain classifier to identify domain-specific SAE features:

bash scripts/03_run_define_domain_features.sh

---

3. Relapse Classifier Training

Finally, train the downstream relapse classifier while masking the identified domain-specific SAE features:

bash scripts/04_run_train_relapse_classifier.sh

This allows the downstream classifier to operate on an SAE representation with the selected domain-specific features removed.

---

References

Primary Implementation Reference

• cytosae (dynamical-inference/cytosae): The SAE architecture and loss function implementation are adapted from this repository.

Vision Foundation Models

• Bioptimus H0-mini (huggingface.co/bioptimus/H0-mini): Used for extracting embeddings from tissue microarray (TMA) spots

---

License

This project is released under the MIT License.

See LICENSE for the full license text.