Request Access to Polaris™ Breast HRD Small v1

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Polaris™ Breast HRD Small v1 is licensed under CC-BY-NC-ND 4.0 for non-commercial, academic research use only. Additional terms apply — please read the full Terms of Use on this page before submitting. Access is manually reviewed by StratifAI GmbH. By submitting this form you confirm you have read and agree to all terms set out on this page. Request access with your Current and official institutional email (this must match your primary email in your Hugging Face account,other domains will be denied). For research partnerships and collaborations contact [email protected].

Polaris™ Breast HRD Small v1.0

Classification model for homologous-recombination deficiency (HRD) status (MUT vs WT) in breast cancer, predicted directly from H&E whole-slide images. To use Polaris™ Breast HRD Small v1.0 you must first have access to Polaris™FMv2-small. The two models work together, Polaris™FMv2-small processes your slides first, and its outputs are then used by Polaris™ Breast HRD Small v1.0. Open source attribution for this model is provided in Polaris™FMv2-small.

This is the small variant, built on top of the 768-dim Polaris™FMv2-small features (with cat-max pooling). It is a WSI-only model (no clinical covariates).

LICENSE & TERMS OF USE

Please read the Terms of Use carefully before requesting access. By submitting your request you agree to be bound by them.

For clarity, Licensee = You and your institution.

License. This Model and associated code are released under the CC-BY-NC-ND 4.0 license and may only be used for non-commercial, academic research purposes with proper attribution. Any commercial use, sale, or other monetization of the Model and its derivatives, which include models trained on outputs from the Model or datasets created from the Model, is prohibited and requires prior approval. By downloading this model, you agree not to distribute, publish or reproduce a copy of the model. If another user within your organization wishes to use the Model, they must register as an individual user and agree to comply with the terms of use. Users may not attempt to re-identify the deidentified data used to develop the underlying Model.

No Ownership Transfer. This license does not constitute a transfer of ownership. StratifAI GmbH retains full ownership of the Model, its weights, architecture, all associated Intellectual Property, and related Know-How. The underlying DINOv2 and VIRCHOW components remain subject to their terms and the Apache 2.0 license and are the property of Meta Platforms, Inc.

Findings. As part of our research collaboration, we request that you share a summary of your findings with StratifAI before any public disclosure or publication. This helps us improve the Model and support your work.

Primary Use Cases

Validation studies with vetted research and clinical partners
Benchmarking and performance evaluation in computational pathology

Out-of-Scope Uses

Clinical deployment without regulatory approval
Any commercial use without a separate written agreement

DISCLAIMERS & WARRANTIES

The Model is provided "AS IS" without warranties of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, or non-infringement. The Licensee acknowledges that the Model has not yet been approved by regulatory authorities for any clinical use, and that it is provided without any warranty or guaranteed availability. The Model is available for Research Use Only (RUO).
Model outputs do not constitute professional or medical advice, including as a substitute for clinical judgment, diagnosis, treatment, or prevention of disease.

REVOCATION OF ACCESS

Revocation for Breach. StratifAI GmbH reserves the right to immediately revoke access to the Model, without prior notice, in the event of any breach of these terms

Contact

Purpose	Contact
Access & research partnerships	[email protected]
Legal & licensing enquiries	[email protected]
Press & marketing	[email protected]

Model details

Target: HRD_Binary — MUT vs WT (classification)
Feature extractor: Polaris™FMv2-small with cat-max pooling (StratifAI/PolarisFMv2-small, 768-dim, tile_size_px=224, tile_size_um=224)
Architecture: masked_vision_transformer (WSI-only, no clinical covariates)
Output classes: ["MUT", "WT"]

Performance

5-fold cross-validation on TCGA-BRCA (held-out validation AUROC per fold):

fold	val AUROC
0	0.718
1	0.833
2	0.798
3	0.756
4	0.755
CV mean	0.772

The model sits on a ~0.78 CV-AUROC plateau: an extensive sweep of learning rate, dropout, weight decay, and feature augmentation produced no improvement beyond noise once averaged over seeds and folds (the ~199-patient validation split has a ≈±0.03 per-run AUROC noise floor). The packaged weights are the fold-3 checkpoint (val AUROC 0.756), the fold closest to the cross-validated mean.

Intended use & limitations

Subtype scope: performs in HR+/HER2− and HER2+; poor on TNBC — do not rely on it there.
Not externally validated. All reported numbers are internal TCGA-BRCA cross-validation; independent external validation is pending and required before any prognostic/clinical claim.
Research Use Only — see Disclaimers above.

Usage

This model is available on HuggingFace:

sai-predict \
    --model-url https://huggingface.co/StratifAI/polaris-hrd \
    --feature-dir /path/to/polarisfm-small-cat-max-features \
    --slide-csv /path/to/slide.csv \
    --output ./predictions.json

Or using local config/weights:

sai-predict \
    --model-config ./config.json \
    --weights ./weights.pt \
    --feature-dir /path/to/polarisfm-small-cat-max-features \
    --slide-csv /path/to/slide.csv \
    --output ./predictions.json

Features must be extracted with StratifAI/PolarisFMv2-small using cat-max pooling (768-dim, tile_size_px=224, tile_size_um=224). This is a WSI-only model — no clinical covariates are required.

How this model was created

Source

Trained via the MCP train_classification_url_ pipeline (sai-engine-train), config crossval_wsi_only, using Polaris™FMv2-small (224px/224µm, cat-max pooling) features.

Data

Cohort: TCGA-BRCA, WSI-only.
Labels: HRD_Binary (Markowetz HRD variable; binary), n = 167 MUT / 431 WT.
Split: 5-fold cross-validation, ~597 train / ~199 validation patients per fold.

Model architecture

masked_vision_transformer v1.0 — dim_input=768, dim_model=384, mlp_dim=1536, n_heads=8, n_layers=2, dropout=0.25, 2-class output. No clinical covariates.

Checkpoint selection

Packaged weights = the top-k-by-auroc_valid_window checkpoint of fold-3 (step 1050, valid AUROC 0.756). Fold-3 was chosen as the single representative model because its validation AUROC is closest to the cross-validated mean (0.772). Selection on this dataset uses multi-seed, paired-per-fold CV — single-fold ranking is unreliable given the ≈±0.03 noise floor.

Model card last updated: 2026-06-16 — Polaris™ Breast HRD Small v1 StratifAI GmbH Karl-Liebknecht-Str. 29A, 10178 Berlin, Germany stratifai.com

Downloads last month: -