Expert-grounded inspection of sparse autoencoder features in histopathology foundation models.
Paper | PDF | Poster | SAE checkpoint | Evaluation protocol | SAE configuration | Example panels | Citation
Histoscope is an interactive dashboard for examining sparse autoencoder (SAE) features learned from pathology foundation-model embeddings. It connects feature activations to tissue classes and to the image patches that activate each feature most strongly, allowing domain experts to inspect whether a feature represents a coherent morphological concept.
The accompanying study used a fixed TopK SAE trained on UNI embeddings from SPIDER-colorectal and a blinded two-pathologist review of 100 confidence-stratified features. Histoscope's AUPRC rule precisely identified class-concentrated features, but missed coherent morphology that appeared across diagnostic classes. The paper therefore recommends using class selectivity to prioritize inspection rather than to remove low-selectivity features from view.
| Component | Location | Notes |
|---|---|---|
| Histoscope dashboard | histoscope/ |
Heatmap, patch browser, feature, cluster, flipper, and report views |
| Study applications | apps/ |
Vocabulary-builder and blinded rating interfaces |
| SAE and artifact pipeline | pipeline/ |
Train the SAE and generate the caches consumed by Histoscope |
| SAE configuration | configs/ |
Exact architecture and training configuration used in the paper |
| Evaluation protocol | protocol/ |
Vocabulary, rater instructions, thresholds, and feature-sampling procedure |
| Reviewed examples | examples/ |
Real top-activating patch panels used for qualitative illustration |
| Workshop paper | paper/ |
Camera-ready ICML Mechanistic Interpretability Workshop paper |
| Checkpoint and caches | models/README.md |
Verified checkpoint on Hugging Face; generated caches remain local |
The release does not contain raw rater responses or the 100-feature expert-label dataset. Those annotations form part of a later extended study. SPIDER-colorectal images and UNI weights are governed by their upstream gated-access terms and are not redistributed here.
Histoscope requires the released SAE checkpoint and a generated interactive cache described in models/README.md.
git clone https://github.com/mnhcorp/histoscope.git
cd histoscope
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python histoscope.py --models-dir /path/to/sae-modelsOpen http://localhost:8050. Image paths recorded in the interactive cache must resolve locally for patch views to render.
Optional patch descriptions use the Gemini API when GEMINI_API_KEY is set. MedGemma support additionally requires transformers and can be enabled with --medgemma; neither integration is required for the dashboard's core inspection workflow.
Histoscope is the final visualization stage, not the artifact generator. The complete path is:
cached UNI embeddings -> TopK SAE training -> feature analysis/cache export -> Histoscope
pipeline/run_pipeline.py chains training and analysis using the paper configuration. It can also skip training and build the dashboard artifacts from the released checkpoint. Exact commands and input layout are documented in pipeline/README.md.
The paper's SAE maps 1,024-dimensional UNI embeddings to a 49,152-feature dictionary using tied encoder/decoder weights, a learned pre-encoder bias, and per-patch TopK activation with k=250. It was trained for two epochs with batch size 32, learning rate 1e-4, seed 42, z-score normalization, and row-wise L2 normalization. Test embeddings use training-split normalization statistics in the corrected release pipeline.
The machine-readable receipt is configs/uni_spider_topk250.json. Download and checksum verification instructions are in models/README.md.
The released protocol includes:
- the controlled morphology vocabulary and its seven top-level families;
- rater-facing instructions and rating scales;
- the prespecified AUPRC, margin, and recall thresholds;
- the confidence-stratified feature-sampling procedure; and
- the Gradio applications used to construct the vocabulary and collect blinded ratings.
Start with protocol/EVALUATION_PROTOCOL.md. This protocol is provided for replication and extension; it should not be interpreted as a release of the original expert annotations.
histoscope/ Dash dashboard
apps/ Gradio vocabulary and rating applications
pipeline/ SAE training and Histoscope artifact generation
configs/ Model and preprocessing receipts
protocol/ Evaluation protocol and controlled vocabulary
examples/ Real reviewed-feature patch panels
models/ External checkpoint/cache instructions
paper/ Workshop paper
@inproceedings{hossain2026histoscope,
title = {Histoscope: Expert-Grounded Inspection of Sparse Autoencoder Features in Histopathology Foundation Models},
author = {Hossain, Mirza Nasir and Bell, Sarah L. and Bryson, Gareth and Harris-Birtill, David},
booktitle = {Mechanistic Interpretability Workshop at the 43rd International Conference on Machine Learning},
year = {2026},
url = {https://openreview.net/forum?id=UIqw2I1CQk}
}Histoscope is research software for inspecting learned representations. It is not a medical device and must not be used for clinical diagnosis or decision-making without further validation.
Histoscope source code is released under the GNU General Public License v3.0. Upstream datasets, model weights, and image assets remain subject to their own terms.