Neural additive models for predicting her2 status in breast cancer from immunohistochemical images
Surayuth Pintawong, Thanarat Chalidabhongse, Shanop Shuangshoti · 2024
This study investigated the application of Conformal Prediction (CP) for uncertainty quantifica- tion in Human Epidermal growth factor receptor 2 (HER2) status classification from immunohistochemistry (IHC) images of breast cancer. Although existing machine learning approaches demonstrate high accuracy in predicting HER2 status, they often struggle with borderline cases that require additional testing. We addressed this challenge by implementing a CP framework that generates prediction sets with controlled error rates. Unlike traditional classifiers that provide single predictions, our CP framework can output either singleton predictions for confident cases or multiple labels for uncertain cases, thereby naturally identifying samples that require additional testing. Our methodology employed handcrafted features, including color intensity, local binary patterns, and Haralick features, in combination with tree-based classifiers. The experimental results demonstrated that our approach achieved reliable coverage guarantees across various significance levels, with the XGboost classifier at a significance level of 0.05, achieving accuracies of 80.8% (95% CI: 77.5%–83.4%) and 74.0% (95% CI: 70.6%–77.3%) for positive and negative HER2 statuses, respectively, while reducing the equivocal cases to 69.4% (95% CI: 66.7%–72.1%). The framework provided flexibility in managing the trade-off between the prediction certainty and ambiguous cases. Our analysis revealed that lower significance levels produced more conservative predictions with larger prediction sets, whereas higher levels yielded more selective predictions at the risk of higher misclassification rates. This approach represents a step toward a more reliable and transparent automated HER2 status assessment system.