Knowledge-Based Systems 137 (2017) 19–28
Contents lists available at ScienceDirect
Knowle dge-Base d Systems
journal homepage: www.elsevier.com/locate/knosys
Cell mitosis detection using deep neural networks
Yao Zhou, Hua Mao
∗
, Zhang Yi
Machine Intelligence Laboratory, College of Computer Science, Sichuan University, Chengdu 610065, People’s Republic of China
a r t i c l e i n f o
Article history:
Received 7 March 2017
Revised 15 June 2017
Accepted 24 August 2017
Available online 11 September 2017
Keywords:
Cell mitosis detection
Deep neural networks
Convolutional neural networks
a b s t r a c t
Quantitative analysis of cell mitosis, the process by which cells regenerate, is important in cell biology.
Automatic cell mitosis detection can greatly facilitate the investigation of cell life cycle. However, cell-
type diversity, cell non-rigid deformation and high cell density pose difficulties on handcrafting visual
features for traditional approaches. Aided by massively captured microscopy image sequences, deep neu-
ral networks have recently become available for automatic cell mitosis detection. This paper proposes an
end-to-end framework named as F3D-CNN for mitosis detection, and F3D-CNN is directly trained from
data without requiring designing domain dependent features. Well-trained F3D-CNN first filters out po-
tential mitosis events based on the static information in each individual image, and further discriminates
candidates by incorporating the spatiotemporal information from image sequences. The state-of-the-art
performance of F3D-CNN was confirmed in experiments on two public datasets (multipotent C3H10T1/2
mesenchymal stem cells and C2C12 myoblastic stem cells).
©2017 Elsevier B.V. All rights reserved.
1. Introduction
Cell mitosis [1] is a complex process by which mature cells pro-
duce next-generation cells. During this process, the ancestor cell’s
membrane divides to form two new cells, and its genetic mate-
rial is duplicated and evenly distributed. To measure cell prolifer-
ation and analyze the cells’ responses to various stimuli, cell bi-
ologists usually perform tedious and time-consuming procedures
in wet laboratories. In particular, they monitor cells over time to
collect informative data, then study the cell dynamics. However,
modern microscopy image capture systems can automatically and
regularly take images of the monitored cells [2] . Using computer
vision based approaches, cell mitosis can be studied from a large
volume of collected high-quality biomedical data without interven-
ing with cell processes [3] . Apparently, there is a keen require-
ment for automatic and robust approaches that can detect the time
and location of cell mitosis events from given image sequences [1] .
As cells undergo non-rigid deformations, and are generally diverse
and densely packed, developing efficient cell mitosis detection ap-
proaches remains a challenging problem.
Deep neural networks (DNNs) have achieved state-of-the-art
performance in various tasks [4,5] , as they can automatically learn
representative features from high-dimensional data [6] . With rep-
resentation learning [7] , the performance of data-driven mitosis
∗
Corresponding author.
E-mail addresses: zy3381@gmail.com (Y. Zhou), huamao@scu.edu.cn (H. Mao),
zhangyi@scu.edu.cn (Z. Yi).
detection from histology images has been improved [8,9] . Convo-
lutional neural networks (CNNs) [10] , which constitute one class of
DNNs, differ from traditional multilayer perceptrons (MLPs) by em-
ploying local connectivity and shared weights to reduce the num-
ber of free parameters, thereby preventing over-fitting problems.
In microscopy images, modeling spatiotemporal features are im-
portant for mitosis detection rather than only focus on static fea-
tures [1,11,12] . In 3D convolutional neural networks (3D-CNNs), the
extended 3D convolutional kernels can process temporal data, e.g.,
human actions can be recognized from image sequences [13] . In
typical CNN-based applications [14] , high-dimensional input im-
ages or image sequences are mapped into (usually) simple re-
sult labels such as classification tasks. Fully convolutional networks
(FCNs) include up-sampling layers that perform image-to-image
prediction [15] . The network output of an FCN can be sized iden-
tically to the input images. CNN and its variants offer several ad-
vantages in cell mitosis detection. First, they automatically learn
robust features from raw data, avoiding the need for domain de-
pendent feature designing. Second, 3D-CNNs can efficiently capture
both spatial and temporal features simultaneously. Finally, CNNs
can be easily parallelized on computing platforms with graphical
processing units (GPUs) for efficient computing.
In order to automatically detect cell mitosis events from mi-
croscopy image, by combining FCNs and 3D-CNNs, this paper pro-
poses a deep neural network named as F3D-CNN. F3D-CNN com-
prises two stages: candidate detection and mitosis discrimination.
In the candidate detection stage, after learning static features of
cell mitosis events in a supervised manner, FCNs retrieve areas,
http://dx.doi.org/10.1016/j.knosys.2017.08.016
0950-7051/© 2017 Elsevier B.V. All rights reserved.