HEp-2 cells Classification via clustered multi-task learning
Anan Liu, Yao Lu, Weizhi Nie
n
, Yuting Su
n
, Zhaoxuan Yang
School of Electronic Information Engineering, Tianjin University, Tianjin 300072, China
article info
Article history:
Received 28 February 2015
Received in revised form
13 May 2015
Accepted 1 June 2015
Available online 22 February 2016
Keywords:
HEp-2 cell
Clustered multi-task learning
Cell classification
abstract
This paper proposes a clustered multi-task learning-based method for automated HEp-2 cells Classifi-
cation. First, the visual feature is extracted for individual sample to represent its appearance char-
acteristics. Then, the models of multiple HEp-2 cell category are jointly trained in the framework of
clustered multi-task learning. The extensive experiments on the HEp 2, cell dataset released by the HEp-
2 Cells Classi fication contest, held at the 2012 International Conference on Patter Recognition, show that
the proposed method can discover and share the latent relatedness among multiple tasks and conse-
quently augment the performance. The quantitative comparison against the state-of-the-art methods
demonstrates the superiority of the proposed method.
& 2016 Elsevier B.V. All rights reserved.
1. Introduction
Over the last few years, imag e processing and pattern recognition
techniq ues hav e been w idely lev erag ed to develop the computer-
aided diagnosis (CAD) systems. Although these systems canno t make
the exact decision, they can work as a pre-selection of the cases for
further examining and conseq uently enable the ph ysici an to focus the
attention only on the most relev a nt cases [1–4]. Therefore, humans
hav e applied these systems into multiple fields of life science and
medical science both for research purposes and for actual clinical
practi ce [5].
Recently, more and more researchers are paying attention on
developing CAD syst ems to realize automat ed analysis of indirect
immunofluorescence (IIF) images. IIF is a diagnostic methodology
based on image analysis that reveals the presence of autoimmune
diseases by searching for antibodies in patient's serum [5].However,
physicians can only treat IIF with specific subjectiv e methods which
highly rely on the experience and expertise. This has caused sig-
nificant disagreement for further diagnosis. It has been reported that
the inter-laboratories agreement is 92.6% for the simple task of posi-
tive/negativ e intensity classification, while it can drop to 7 6 .0% for the
recogni tion of staining patter ns, which is req uir ed for a mor e detailed
diagnosis [6]. Therefore, in recent years, lots of work have been done
on the related research topics, including image preconditioning [7,8],
image segmentation [9,10], mitotic cell recognition [11–17], and pat-
tern recognition [18–20 ]. Since the success of HEp-2 Cells Classifica-
tion contest held at 2012 International conference on Patt ern R ecog-
nition, more and more researchers are being engaged in this task. The
current methods of HEp-2 cell classification usually contain three
main steps, including feature extraction, feature selection, and model
learning. For feature e xtraction, Cheplygina et al. utilized the variance
and covariance of intensity values, the histograms of the red and green
channels, and the morphological features of the foreground as visual
representation [5]. T o improve the discrimination of feature repre-
sentation, many textual featur es, such as local binary patterns [21],
gray level co-occurrence matrix [22], and discrete cosine transform
coeffici ents, and sophis ticat ed shape features, such as hist ograms of
oriented gradients [23], have been wide leveraged for this task. Since
much work extracted multiple features for representation, feature
selection is necessary. Kazanov et al. used forward selection to extract
the significant features from the basic intensity features and mor -
phological features [5]. Mateos-Garcia et al. leveraged the correlation
feature selection and genetic algorithm for feature selection
[5].
Rezv an i et al. utilized the classic PCA to select the best features from
the ex tracted textua l and intensity features [5].Withthedis-
criminative features, lots of powerful classifiers are implemented for
cell classification, including k- NN, support vector machine (SVM)
[24–26], multiclass SVM [27], ShareBoost [28], AdaBoost [25],and
random decision forest [29].
Although much work has been done for HEp-2 cell classificatio n,
there still exist two critical problems. To our knowledge, most of the
previous methods work in the framew ork of single-task learning.
Although man y sophisticated classifiers have been designed, they
cannot take the latent relatedness among multiple cell categori es into
considera tion. It is challenging to cluster HEp-2 cells into clusters
simply with the visual appearances by humans. Therefore, it is extr e-
mely difficult to discriminate several clusters (such as homogeneous,
fine speckled, and coarse speckled) with similar patterns. On the other
hand, there usually exist limited samples since it is expensive for
large-scale dataset preparation and manual segmentation and
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/neucom
Neurocomputing
http://dx.doi.org/10.1016/j.neucom.2015.06.108
0925-2312/& 2016 Elsevier B.V. All rights reserved.
n
Corresponding authors.
E-mail addresses: weizhinie@tju.edu.cn (W. Nie), ytsu@tju.edu.cn (Y. Su).
Neurocomputing 195 (2016) 195–201
下载后可阅读完整内容，剩余6页未读，立即下载
weixin_38701683
- 粉丝: 4
- 资源: 927
我的内容管理
展开
最新资源
- 李兴华Java基础教程：从入门到精通
- U盘与硬盘启动安装教程：从菜鸟到专家
- C++面试宝典：动态内存管理与继承解析
- C++ STL源码深度解析：专家级剖析与关键技术
- C/C++调用DOS命令实战指南
- 神经网络补偿的多传感器航迹融合技术
- GIS中的大地坐标系与椭球体解析
- 海思Hi3515 H.264编解码处理器用户手册
- Oracle基础练习题与解答
- 谷歌地球3D建筑筛选新流程详解
- CFO与CIO携手：数据管理与企业增值的战略
- Eclipse IDE基础教程：从入门到精通
- Shell脚本专家宝典：全面学习与资源指南
- Tomcat安装指南：附带JDK配置步骤
- NA3003A电子水准仪数据格式解析与转换研究
- 自动化专业英语词汇精华：必备术语集锦
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
