Cell Classi cation from Multi-Focus Images by Deep Neural Network

Ken'ichi Morooka, Jun Ohta, Shoko Miyauchi, Ryo Kurazume, Eiji Ohno

Keywords: Cytodiagnosis, Multi focus images, Deep Neural Network, Imbalanced data

TL;DR: We constructed combined DNN to decect cancer cells automatically for supporting cytodiagnosis.

Abstract:

We present a new method for automatically classi-
fying normal and abnormal cells from multi-focus images by deep neural
networks (DNNs). To construct the DNNs, many images of cancer and
normal cells are needed. However, there is a serious problem of the imbal-
ance between normal and cancer cell images. The data imbalance makes
it difficult to construct the system with acceptable accuracy. To solve this
problem, we develop the cell classication system by combining multiple
DNNs with different architectures using the imbalanced dataset. From
the experimental results, our proposed system can classify cells with high
performance.