首页 关于本刊 编 委 会 期刊动态 作者中心 审者中心 读者中心 下载中心 联系我们 English
 自动化学报  2018, Vol. 44 Issue (10): 1842-1853 PDF
Fisher准则下面向判别性特征的字典学习方法及其组织病理图像分类研究

1. 湘潭大学信息工程学院 湘潭 411105;
2. 湖南大学电气与信息工程学院 长沙 410082;
3. 湘潭大学智能计算与信息处理教育部重点实验室 湘潭 411105

Discriminative Feature-oriented Dictionary Learning Method With Fisher Criterion for Histopathological Image Classification
TANG Hong-Zhong1,2,3, LI Xiao1,3, ZHANG Xiao-Gang2, ZHANG Dong-Bo1,3, WANG Xiang1,3, MAO Li-Zhen1,3
1. College of Information Engineering, Xiangtan University, Xiangtan 411105;
2. College of Electrical and Information Engineering, Hunan University, Changsha 410082;
3. Key Laboratory of Intelligent Computing and Information Processing of Ministry of Education, Xiangtan University, Xiangtan 411105
Manuscript received : December 9, 2016, accepted: September 23, 2017.
Foundation Item: Supported by National Natural Science Foundation of China (61573299, 61673162, 61672216), National Natural Science Foundation of Hunan Province (2017JJ3315, 2017JJ2251, 2016JJ3125), and Scientific Research Project of Hunan Province Education Department (15C1328)
Corresponding author. ZHANG Xiao-Gang  Professor at the College of Electrical and Information Engineering, Hunan University. His research interest covers process control and pattern recognition for industrial kiln. Corresponding author of this paper.
Abstract: The problem of high similarity between learned healthy dictionary and diseased dictionary and low discrimination exists in the current dictionary learning methods for histopathological image feature extraction. In this paper, we present a novel discriminative feature-oriented dictionary learning method based on Fisher criterion (FCDFDL). This method constructs a penalty item of the objective function using Fisher criterion to minimize the intra-class distance of learned dictionaries and maximize the inter-class distance of learned dictionaries. Thus, the similarity between healthy and diseased dictionaries is reduced. Furthermore, the reconstruction of the same class samples is improved over the learned dictionaries, while reconstruction of different class samples is suppressed. Then, the sparse representation of test samples is respectively performed on the learned healthy dictionary and the diseased dictionary, and the classifier is constructed by employing the reconstruction error vector of test samples. Finally, the proposed FCDFDL is tested on ADL and BreaKHis datasets, and experimental results show that the learned dictionaries have stronger discrimination and improved classification performance as compared to the other dictionary learning methods, for histopathological image.
Key words: Histopathological image     Fisher criterion     dictionary learning     discriminative feature

1 DFDL方法

Vu等[29]于2015年提出了一种面向判别性特征的字典学习方法(Discriminative feature-oriented dictionary learning, DFDL), 并应用于医学组织病理图像分类.其目标函数定义如下:

 $\left\langle {D^*, X_D^*, \bar X_D^* } \right\rangle =\notag\\ \qquad \underset{D,{{X}_{D}},{{{\bar{X}}}_{D}}}{\mathop{\arg \min }}\, \bigg( \frac{1}{N}\left\| {Y - DX_D } \right\|_F^2 \, - \notag\\ \qquad \frac{\rho }{{\bar N}}\left\| {\bar Y - D\bar X_D } \right\|_F^2 \bigg)\nonumber \\\, {\rm s.\, t.}~ \left\| {X_D } \right\|_0 < L_1 ~{\rm and}~\left\| {\bar X_D } \right\|_0 < L_1$ (1)
 $\left\langle {\bar D^*, \bar X_{\bar D}^*, X_{\bar D}^* } \right\rangle =\notag\\ \qquad \underset{\bar{D},{{{\bar{X}}}_{{\bar{D}}}},{{X}_{{\bar{D}}}}}{\mathop{\arg \min }}\, \bigg( \frac{1}{{\bar N}}\left\| {\bar Y - \bar D\bar X_{\bar D} } \right\|_F^2\, -\nonumber\\ \qquad \frac{\rho }{N}\left\| {Y - \bar DX_{\bar D} } \right\|_F^2 \bigg)\nonumber\\ \, {\rm s.\, t.}~ \left\| {\bar X_{\bar D} } \right\|_0 < L_2 ~{\rm and}~\left\| {X_{\bar D} } \right\|_0 < L_2$ (2)

3 实验结果及分析

ADL数据集宾夕法尼亚州立大学提供, 包括肺、脾脏、肾脏三类器官, 共计900多张图像.每类器官包括无病和有病两类样本, 各150多张, 尺寸为1 360像素$\times$ 1 024像素.为了提高算法的计算效率, 本文将所有图像归一化为600像素$\times$ 600像素.如图 1所示, 图 1 (a)从左至右依次表示肺、脾脏、肾脏的无病图像, 图 1 (b)从左至右依次表示肺、脾脏、肾脏的有病图像.

 图 1 肺、脾脏、肾脏的组织病理图像 Figure 1 Lung, spleen and kidney images

2) FCDFDL与其他方法的实验对比

3.2 BreaKHis数据集的实验结果

1) BreaKHis数据集及相关实验设置

 图 2 腺病与叶状癌的组织病理图像 Figure 2 The images of adenosis and phyllodes tumor

2) FCDFDL与其他方法的实验对比

3.3 学习字典的类间差异

 图 3 FCDFDL, DFDL, FDDL, LC-KSVD方法学习字典的可视图 Figure 3 The visual maps of learned dictionaries with FCDFDL, DFDL, FDDL, and LC-KSVD method

 图 5 参数$\alpha$, $\beta$的变化对不同病理图像分类精度的影响 Figure 5 Classification accuracy with different parameters $\alpha$, $\beta$ on different pathological images

2) 图块尺寸的设置

 图 6 FCDFDL方法下图块尺寸的变化对不同病理图像分类精度的影响 Figure 6 Classification accuracy on different pathological images with different image block size, and with FCDFDL method
4 总结