Leveraging IHC Staining to Prompt HER2 Status Prediction from HE-Stained Histopathology Whole Slide Images

Yuping Wang, Dongdong Sun, Jun Shi, Wei Wang, Zhi-Guo Jiang, Haibo Wu, Yushan Zheng

Abstract: The development of artificial intelligence has significantly impacted the predictive analysis of molecular biomarkers, which is crucial for targeted cancer therapy. Traditional assessment of HER2 in breast cancer utilizes both Hematoxylin and Eosin (H&E) and Immunohistochemistry (IHC) stained slides. Recent models have sought to predict HER2 status using H&E-stained slides to reduce reliance on the costly and time-consuming IHC staining. However, these models overlook the information from IHC staining. In this paper, we proposes a novel framework that integrates IHC-stained WSIs during the training phase to enhance the HER2 prediction capabilities based on the H&E-stained WSIs. This framework uses IHC-predicted HER2 status as a proxy task, embedding the learned relevant information as prompts into the encoder for H&E slides. Meanwhile, our model only requires H&E slides during inference, which maintains the data-efficiency of the HER2 prediction system. Experimental results show that our method achieves an AUC of 0.860 and a F1 score of 0.652 in the tasks of HER2 0/1+/2+/3+ status grading for breast cancer, which significantly outperforms state-of-the-art models.