Corr-A-Net: Interpretable Attention-Based Correlated Feature Learning framework for predicting of HER2 Score in Breast Cancer from H&E Images

Human epidermal growth factor receptor 2 (HER2) expression is a critical biomarker for assessing breast cancer (BC) severity and guiding targeted anti-HER2 therapies. The standard method for measuring HER2 expression is manual assessment of IHC slides by pathologists, which is both time intensive and prone to inter- and intra-observer variability. To address these challenges, we developed an interpretable deep-learning pipeline with Correlational Attention Neural Network (Corr-A-Net) to predict HER2 score from H&E images. Each prediction was accompanied with a confidence score generated by the surrogate confidence score estimation network trained using incentivized mechanism. The shared correlated representations generated using the attention mechanism of Corr-A-Net achieved the best predictive accuracy of 0.93 and AUC-ROC of 0.98. Additionally, correlated representations demonstrated the highest mean effective confidence (MEC) score of 0.85 indicating robust confidence level estimation for prediction. The Corr-A-Net can have profound implications in facilitating prediction of HER2 status from H&E images.