OpenPink Core (Breast Cancer Histopathology AI)
OpenPink Core is a state-of-the-art Dual-Stream Multiple Instance Learning (MIL) model designed for the classification of Breast Cancer Whole Slide Images (WSIs) and Regions of Interest (ROIs). Built on top of the Virchow2 foundation model, OpenPink classifies tissue into four clinical categories: Atypical, Benign, Malignant, and Normal, with a specific emphasis on identifying diagnostically challenging Atypical lesions.
Model Details
- Architecture: Dual-Stream Gated Attention MIL (processes global macroscopic thumbnails alongside high-resolution 224x224 patches).
- Feature Extractor: Virchow2 (1280-dimensional embeddings, SwiGLU MLP, SiLU).
- Training Strategy: 10-Fold Stratified Group Cross-Validation.
- Domain Adaptation: Multi-source training utilizing Hard Example Mining from South American (BreaKHis) and European (BRACS, BACH) datasets to decouple cellular density from malignancy.
- Ensemble: The repository provides 10 pre-trained checkpoints (folds) for robust ensemble inference.
- License: CC-BY-NC-4.0 (Strictly Non-Commercial & Research use only).
Intended Use & Limitations
FOR ACADEMIC AND RESEARCH PURPOSES ONLY.
- Not for Clinical Use: This tool has not been evaluated by the FDA, EMA, or any other regulatory body. It is not a medical device. It cannot be used for diagnostic, prognostic, or therapeutic purposes.
- Limitations: The model assumes H&E (Hematoxylin and Eosin) stained slides. Variations in extreme staining protocols may degrade performance. We strongly recommend applying Macenko Stain Normalization prior to feature extraction. Note: An automated Macenko normalization pipeline is included directly in the OpenPink-Core repository for your convenience.
Training Data
The model was trained and internally validated on multi-centric European clinical cohorts:
- BRACS (BReAst Carcinoma Subtyping): Hematoxylin & Eosin (H&E) stained ROIs.
- BACH (Breast Cancer Histology): High-resolution microscopy images.
To ensure global generalization, the model was aggressively domain-adapted using "Hard Negative" injections from the BreaKHis dataset (Brazil), focusing on hyper-cellular benign tumors like Fibroadenomas to prevent False Positives in dense tissue structures.
Evaluation Results
Internal Validation (European Cohorts)
During 10-fold CV on BRACS and BACH, the model achieved:
- Accuracy: 93.33%
- Macro F1-Score: 90.18%
- Atypical Recall (Sensitivity): 95.65%
- Malignant Recall: 98.68%
Out-Of-Distribution (OOD) Domain Adaptation
Evaluated on 1,693 unseen images from BreaKHis using a strict Clinical Thresholding constraint (Probability > 0.30 for Malignant).
- Malignant Recall: 82.46% (945 / 1146 cases)
- Benign Specificity: 66.73% (365 / 547 cases) - Massive improvement over the 16.8% baseline.
- Perfect Separation: Zero Atypical or Normal misclassifications in the entire OOD set.
How to Use
The checkpoints in this repository are designed to be loaded by the OpenPink-Core inference scripts.
1. Download Checkpoints via Python
You can automatically download all 10 folds using the huggingface_hub library:
from huggingface_hub import snapshot_download
# Download the checkpoints folder
snapshot_download(
repo_id="MatthewMak/OpenPink-Core",
allow_patterns="*.pt",
local_dir="./checkpoints"
)
2. Feature Extraction & Inference
Clone the OpenPink-Core GitHub/GitLab Repository (link pending) and run the evaluation scripts using the downloaded checkpoints.
# 1. Extract Virchow2 Features (Auto-Macenko enabled)
python extract_features.py --data_dir ./images --output_dir ./features --ref_image ref.png
# 2. Run Ensemble Inference with Clinical Thresholding
python evaluate_ood.py --features_dir ./features --checkpoints_dir ./checkpoints --malignant_threshold 0.30
Citation & Credits
- Author/Lead Researcher: Matthaios Makrogiannis
- Institution: University of Ioannina
- Year: 2026