This repository contains the implementation code for the paper Conformal Prediction for Uncertainty Quantification and Reliable HER2 Status Classification in Breast Cancer IHC Images.
This project implements a Conformal Prediction (CP) framework for HER2 status classification in breast cancer IHC images. The framework:
- Quantifies uncertainty in HER2 status predictions
- Identifies borderline cases requiring additional testing
- Provides prediction sets with controlled error rates
- Uses handcrafted features with tree-based classifiers
- Python installed on your system
- At least 10GB free disk space (dataset: 6GB + extracted features)
- Clone repository:
git clone https://github.com/Surayuth/her2cp.git
cd her2cp- Create and activate Python virtual environment:
python -m venv .venv
source .venv/bin/activate # For Unix/macOS
# Or for Windows:
# .venv\Scripts\activate- Install dependencies:
pip install -r requirements.txt- Download dataset (6GB):
- Download dataset file
- Save to the project root directory
- Extract dataset:
unzip Data_Chula.zip- Extract features (estimated time: 1-2 minutes):
python extract_feature.py --workers 8 --src ./Data_Chula --dst ./extracted_featuresExample training command (using Decision Tree, completes in ~15 minutes):
python train.py --path extracted_features/feat_level_16_scale_0.25.csv --model dtAvailable models: Decision Tree (dt), Random Forest (rf), Gradient Boosting Tree (gbt), and Extreme Gradient Boosting Tree (xgb). Note: Training time varies by model complexity.
The following notebooks reproduce the key results from the paper:
-
Miscoverage Rate Analysis:
Analysis/miscoverate.ipynb- Verifies the framework's coverage guarantees at different significance levels
-
Ambiguity Rate Analysis:
Analysis/ambiguity.ipynb- Analyzes the trade-off between prediction certainty and ambiguous cases
-
Accuracy Analysis:
Analysis/accuracy.ipynb- Evaluates classification performance for positive and negative HER2 statuses
- All figures are standardized with y-axis limits set to [0, 100]
- The main differences between notebooks are in the aggregate functions used in the 4th cell