Conformal Prediction for Uncertainty Quantification and Reliable HER2 Status Classification in Breast Cancer IHC Images

Accurate assessment of Human Epidermal growth factor receptor 2 (HER2) status in breast cancer is crucial for determining treatment eligibility for targeted therapies. Although machine learning approaches have shown promise in automating HER2 status classification from immunohistochemistry (IHC) images, they struggle to maintain reliable accuracy in equivocal cases. This study introduces a novel application of Conformal Prediction (CP) framework that, instead of making potentially incorrect predictions for uncertain cases, explicitly identifies them for further examination. The CP framework achieves this by generating prediction sets whose size reflects uncertainty, inherently aligning with the clinical practice of identifying cases that require confirmatory testing. Using dual in situ hybridization (DISH) results as the gold standard for HER2 status, we developed our system using handcrafted features—including color intensity, local binary patterns, and Haralick features—combined with tree-based classifiers. In experiments on 2,570 IHC images (94 cases) from the Institute of Pathology, Bangkok, Thailand, our framework demonstrated the ability to make highly accurate predictions on unambiguous cases while automatically filtering out uncertain ones. At a significance level of 0.05, the XGBoost-based conformal predictor achieved 90.1% (95% CI: 87.4-92.2%) accuracy on the equivocal cases, while identifying 69.8% (95% CI: 67.7-71.9%) of cases as requiring additional testing. This selective prediction approach represents a practical step toward reliable automated HER2 assessment by focusing on high-confidence predictions. The source code of our implementation is available at: https://github.com/Surayuth/her2cp.git.