A Comprehensive Study on Classification of Breast Cancer Histopathological Images: Binary Versus Multi-Category and Magnification-Specific Versus Magnification-Independent

Abstract

There are millions of cancer cases worldwide every year, and breast cancer is one of the most prevalent diseases with the highest mortality rate. The manual effort of pathologists can be significantly reduced by computerized diagnostic systems, which improve the accuracy and reliability of diagnosis. In this paper, we present four novel systems for breast cancer diagnosis in four different scenarios: binary versus multi-class classification and magnification-specific (MS) versus magnification-independent (MI) classification. In each of the proposed systems, we developed an automatic score-level fused CNN model using a pretrained deep neural network and named it the Multi-Level Feature Fusion (MLF2) model. The MLF2-CNN, similar to the conventional CNN models, integrates the feature extraction and classification phases of BC classification into a single automatic learning procedure. Additionally, MLF2-CNN performs an automatic score-level fusion of several classifiers that were trained with multi-level features to make the final decision. A pretrained DenseNet-121 is selected as the backbone of the proposed MLF2-CNN, and several new links are added to the CNN architecture to capture multi-stage features. Several experiments on the publicly available BreakHis dataset demonstrate that the proposed systems capture the best descriptive features and outperform state-of-the-art techniques in most of the scenarios.

Keywords

• Convolutional neural network
• breast cancer
• BreakHis dataset
• histopathological images
• score-level fusion
• directed acyclic graph