Projects Last Projects

HER2 status prediction from H&E-stained WSI by utilizing HER2-IHC images and automatic tile level annotations in breast cancer.

Amit Frechter

Supervised by Gil Shamai and Shachar Cohen

Abstract

Accurate and efficient Her2 detection from Hematoxylin and Eosin (H&E) stains remains challenging and crucial for treatment decisions. This project suggests a new method that utilizes Immunohistochemistry (IHC) images and patch-level
annotations. In this study several deep learning models were developed to try and exploit H&E - IHC whole slide images (WSI) pairs to improve the Her2 prediction capabilities from H&E stained images.
One method included producing patch-level score matrices to enable the model to learn local scores that would later be aggregated to a slide-level score. The other method used the Teacher-Student (knowledge distillation) approach to train a model from H&E images that imitates the behavior of a stronger model
that was trained on IHC images.

Pictures

Project HER2 status prediction from H&E-stained WSI by utilizing HER2-IHC images and automatic tile level annotations in breast cancer. Picture 1