1226P - Multi-site validation of a deep learning solution for HER2 profiling of breast cancer from H&E-stained pathology slides

Background

Molecular profiling of the human epidermal growth factor receptor 2 (HER2) is performed for all malignant breast cancers to inform the choice of HER2-targeted therapy. IHC and ISH are performed to establish HER2 status in standard clinical care. These tests are not only expensive and time-consuming, but also account for most of the turnaround time for diagnosis and require expert interpretation. Here, we demonstrate the effectiveness of PANProfiler Breast (HER2 Negative), a UKCA-marked deep learning (DL)-based image analysis software for HER2 profiling from whole slide images (WSIs) of routinely-used H&E-stained pathology slides.

Methods

A DL model was trained and validated on 1684 WSIs from two datasets to identify HER2-negative cases defined by IHC0, IHC1 or IHC2+/ISH-. A total of 2310 H&E images of breast cancer samples from five different sites in the UK and US were used for external validation. These images were acquired with four different scanners with each set containing a varying proportion of biopsies and resections. The model was evaluated separately for each site with 3-fold cross-validation and results were aggregated across folds. The performance was measured in comparison to HER2 status acquired via IHC and ISH in accordance with RCPath & CAP guidelines.

Results

Incidence rate for HER2-positive cases varied between 9-15% in the UK sites and was 64% for the US cohort. Overall accuracy across all sites was 93.06% (± 8.28%), reaching up to 97.99%. An average false negative rate (discordance to wet-lab assay) of 9.27% (± 0.47%) was achieved with a 100% specificity for all datasets. Performance was robust to specimen/scanner types, with accuracies of 96.46%, 95.39%, and 91.15% achieved in sites that contained only biopsies, only resections, and mixed types. Accuracy across scanners varied by a standard deviation of 9.25%.

Conclusions

We demonstrate the robustness of a DL-based HER2 profiling method in breast cancer utilising only H&E-stained WSIs. This multi-site validation study is the first-of its kind for such an approach using real-world clinical data. Our solution could facilitate fast, accurate, and systemic screening of patients for targeted treatments if integrated within the routine pathological workflows.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Panakeia Technologies Ltd., NHS, Innovate UK.

Disclosure

All authors have declared no conflicts of interest.