Predicting HER2 status in esophageal cancer now on bioarxiv

We show how to automatically predict HER2 score in immunohistochemically stained slides of esophageal adenocarcinoma.

HER2 is an important cancer marker that in esophageal carcinoma allows the patients to receive a targeted treatment. To identify HER2-positive samples pathologists use immunohistochemical (IHC) staining of HER2, which does not allow to recognize the HER2 status in all patient cases. We developed a weakly supervised classification approach with attention mechanism to determine the HER2 status in the IHC images. In our test set of ~350 images we achieve 0.94 accuracy in prediction of HER2 status. We offer visual insights into the variability of HER2 activity across tumors (such as in the figure shown here). We also demonstrate how attention mechanism can allow us to identify image regions that are key for making the prediction.

Find out more in our preprint and visit github for code and tutorials.