CNN-Based Mitosis Detection for Assisting Doctors to Diagnosis

Abstract

Breast cancer kills more than 500,000 people every year all over the world. The number of mitotic cells is one important indicator to evaluate breast cancer progressing, and mitotic counting is very tedious and easy to make mistake. Therefore, the development of computer-aided detection (CAD) system is important in diagnosis and treatment of breast cancer. In this paper, we propose a CAD system based on a convolutional neural network (CNN) to automatically count mitotic cells. The proposed system consists of three steps. First, a normalization process is exploited to reduce the illumination variance and noise among different individuals as well as highlight the nuclei regions, which can simplify the problem of mitotic counting. Second, an improved convolutional neural network based on LeN et-5 is established to extract features. Last, Softmax can finally determine the location of mitotic cells. Our CAD system was evaluated on the dataset provide by 2012 ICPR mitosis detection challenge, and experimental results revealed that F1 score achieved 0.884.