HER2 Status Prediction in Breast Cancer: A Weak Supervision and Contrastive Learning Approach on H&E Stains

In order to treat breast cancer patients with anti-HER2 therapies, their HER2 status is determined using immunohistochemistry (IHC). Cases that are equivocal by IHC (grade 2+) are sent for fluorescence in situ hybridization (FISH). FISH can be expensive and cause treatment delays for IHC-equivocal cases, but hematoxy lin and eosin (H&E) staining of tumor tissues is widely available. We developed a customized pipeline for self-supervised feature extraction and weakly supervised whole slide image (WSI) classification of H&E stained tissue. We trained our model on publicly available WSIs from The Cancer Genome Atlas (TCGA), with annotations from the Yale School of Medicine. Our pipeline achieved an AUC of 0.87 ± 0.05 across four test folds. We additionally tested the model on 44 difficult held-out cases that were equivocal by IHC and achieved an AUC of 0.80 against FISH status. This approach could help reduce reliance on expensive FISH tests and improve cancer treatment equity for under-served populations.