Computer-Aided Breast Cancer Diagnosis Using Deep Learning: Malignancy Detection and HER2 Scoring

Breast cancer is the leading cause of cancer related deaths in women worldwide. Proper diagnosis is highly essential to determine the appropriate patient treatment based on their specific tumor type. Human epidermal growth factor receptor 2 (HER2) gene amplification is considered a poor prognostic factor in breast cancer and an excellent predictive marker. Fortunately, breast cancer survival rates were shown to significantly improve upon following anti-HER2 targeted therapies. Diagnosis of breast cancer has many phases: clinical, radiological, and pathological diagnosis. The role of artificial intelligence (AI) within radiological diagnosis is now well established and accepted by medical practitioners. In this study, a two-phase deep learning-based diagnostic tool is presented to aid in the pathological diagnosis of breast cancer. In the first phase, breast cancer classification (benign vs. malignant) was performed using hematoxylin & eosin (HE) images. In the second phase, HER2 classification (positive vs. negative) was considered using immunohistochemistry (IHC) images. Four pretrained networks with different architectures were compared in both phases which are EfficentNetB0, InceptionV3, ResNet34, and VGG19. For cancer detection, InceptionV3 and VGG19 gave the highest accuracy of 97% using the public BreakHis dataset. As for HER2 classification, an accuracy of 100
^%
was achieved by the four considered networks for a private dataset. This work lays the foundation for creating a reliable and comprehensive automated screening tool that would assist in the early detection and proper treatment of breast cancer.