Magnification-specific and magnification-independent classification of breast cancer histopathological image using deep learning approaches

Breast cancer (BC) is a massive health problem and a deadly disease, killing millions of people every year. Computerized approaches for automated malignant BC detection can efficiently help in reducing the manual workload of pathologists and making diagnosis more scalable and less prone to errors. In this paper, we present two systems to diagnose breast cancer from single and multi-magnification histopathological images. The first proposed system utilizes a pre-trained DenseNet201 CNN architecture and fine-tuned over the publicly available BreakHis dataset and classifies histopathological images of specific magnification factors into one of the benign or malignant classes. The second system consists of four subsystems, each corresponding to one of the magnifications, and is trained only by related magnification images. Afterwards, the results obtained from these four subsystems are fused together to make the final decision. Several experiments on BreakHis dataset demonstrate that the proposed systems outperform the state-of-the-art approaches, in all cases.