HER2-SISH HISTOPATHOLOGY IMAGE CLASSIFICATION USING DEEP NEURAL NETWORKS

The status of the human epidermal growth factor receptor 2 (HER2) gene amplification is an important marker for assessing the efficacy of clinical treatments for breast cancer. This article discusses the application of deep learning to classify HER2-SISH (silver-enhanced in situ hybridization) pathological images and identifies their HER2/Chr17 amplification status. We used four pre-trained models for classifying the cases into either amplified or non-amplified: two models from the convolutional neural networks, CNNs (DenseNet, and MobileNet), and two transformer models (Vision Transformer, and Data-Efficient Image Transformers). Apart from these single models, we also built two ensemble models by concatenating the transformer and CNN architectures to observe their performances. A private dataset obtained from our collaborating hospital is used in this project, with several preprocessing techniques applied to the raw images prior to feeding the models. Promising results are reported with ViT emerged as the best performing model with a high accuracy of 87.47%, with 92.93% sensitivity in detected amplified HER2-SISH samples. This paper is of high novelty due to the limited literature found working on the newly introduced stain and this is the first work using data-driven approach for the HER2-SISH biomarker.