A Robust HER2 Neural Network Classification Algorithm Using Biomarker-Specific Feature Descriptors

Computer assisted evaluations of Whole Slide Images (WSI) of histopathological slides require robust biomarker-specific feature descriptors for accurate grading and classification. Considering the large amount of processing involved in analysing WSIs, training and classification, it is important to have an optimized set of features that closely represent the characteristics of the biomarkers used by pathologists in manual assessments. In this paper, we consider the problem of classifying WSIs of ImmunoHistoChemistry (IHC) stained slides for automated breast cancer grading. We use a combination of intensity and texture features derived from the input image at different saturation levels, and show its effectiveness in a Neural Network architecture for classifying the image into one of the four HER2 scores. The paper also presents three configurations for the neural network and gives comparative analysis showing the variations of classification accuracy with respect to changes in the configuration and the learning rate.