Challenge
Manual HER2 immunohistochemistry (IHC) scoring has high variability, especially for HER2 ‘low’ (1+) and HER2 ‘ultralow’ (between 0 and 1+) categories characterizing rare staining, which could impact breast cancer treatment decisions, particularly with respect to antibody-drug conjugates (ADCs).
A leading cancer non-profit organization is running a multi-institutional digital pathology HER2 harmonization study using samples from breast cancer patients to improve consistency in patient treatment decisions.
Solution
The Nucleai team was chosen as one of the algorithm developers chosen this effort. We scored >2000 paired HER2 IHC and hematoxylin and eosin (H&E) slides within 4 weeks with a locked model. A locked model, often used for testing and validation, is defined as one whose parameters do not change after the model has been trained and tested. A subset of the HER2 slides were also scored by an expert pathologist.
We found that the AI model found rare staining (>1% positive cells) in ~63% of the HER2 IHC slides scored as negative by manual pathologist review, suggesting that these samples could be potentially recategorized as HER2 ultralow, typically with a score between 0 and 1+ (Figure 1).
This study highlights the limitations of manual HER2 scoring and the rare HER2+ cells that could be missed by the human eye (Figure 2). AI-aided scoring could expand the dynamic range of detection to low expressors, enabling the inclusion of additional patients that can be missed by manual scoring.
Figure 2: An example of a patient sample scored as HER2 negative with manual pathologist review. The AI model noted ~2.41% HER2 positive cells in the sample which suggests it could be potentially recategorized as HER2 ultralow.
Reference:
- Mehta, Sandhya, et al. “Prevalence of ‘HER2 Ultra-Low’ among with Advanced Breast Cancer with Historical IHC0 Status.” Journal of Clinical Oncology, vol. 42, no. 16_suppl, June 2024, pp. e13156–e13156