Machine Learning Models to Quantify HER2 for Real-Time Tissue Image Analysis in Prospective Clinical Trials
HER2 overexpression is a demonstrated negative prognostic factor in breast cancer, and a target for anti-HER2 compounds. There is an unmet need for reproducible and accurate HER2 scoring in breast cancer as it is essential to inform treatment decisions. Pathologists show inter- and intra-observer variability for whole slide quantitative scores in part because exhaustive cell scoring is not possible by manual means. Using machine learning (ML) approaches, every viable tumor cell within the HER2 stained sample is classified. Here, we developed, trained, and validated an automated ML-based model as a quality control tool for HER2 testing and monitoring in clinical trials. The ML model was trained using whole slide images (WSI) from multiple sources to quantify HER2 expression, and measure stain intensity, artifact content, tumor area, and DCIS (ductal carcinoma in-situ) across a diversity of breast cancer phenotypes. Model quantified HER2 scores were consistent with pathologist consensus scores across breast cancer tissue types. These results support incorporation of ML-based algorithms into clinical trial workflows to monitor HER2 testing quality including scoring, tissue quality, and assay performance.
Glass et al.