HEp-2细胞间接荧光免疫图像识别方法研究

This project is focused on some issues related to staining pattern classification for HEp-2 indirect immunofluorescence cell images. These issues mainly include methods to accurately locate boundaries of HEp-2 cells, methods to extract discriminative texture features and fusion methods of different kinds of features. Particularly, a coarse-to-fine framework is proposed to locate the boundaries of cells. The rough cell regions are located first and then refined by some level set or GVF Snake based methods. To extract robust while discriminative features, several up-to-date feature extraction methods are investigated and improved according to the appearance of HEp-2 cell images. Two new texture descriptor, namely maximum entropy based local multiple pattern and block based Gabor coefficients statistic, are proposed. The compressed sensing and sparse encoding techniques are also analyzed to introduce some new texture feature extraction methods suitable for HEp-2 cell images. Finally, the feature fusion methods are examined at both the feature level and decision level. MCCA is used to remove the redundance between diffenrent features, trying to get more accurate classification results. A feature fusion framework based on posteriori probability classifier and AdaBoost.M1 framework is also introduced. The system which adopts some of the proposed methods above attented the HEp-2 cell staining pattern classification contest in ICPR2012,and got quite good results. In a word, this project is bound to have wide application prospects, and is of practical significance to improve the clinic level and to ensure the health of public.

本项目以HEp-2细胞间接免疫图像荧光型别分类为主要研究内容，结合当前的研究现状与发展趋势，贯穿整幅图像分割、细胞特征提取及多特征融合分类三个方面。提出由粗到细的分割思路，粗粒度层面：提取细胞区域，细粒度层面：精确定位细胞边缘；探索鲁棒且区分能力强的特征提取方式，首先在分析现有纹理提取方法问题基础上，结合HEp-2细胞图像特点提出改进方案，提出最大熵多值模式和分块Gabor系数加工统计思路；其次，结合压缩感知和稀疏表示，探寻适合HEp-2细胞图像的纹理提取方法；对多种特征进行融合分类，特征层面上，使用MCCA消除不同特征间的冗余信息；决策层面上，提出基于后验概率的融合分类框架。本项目前期研究成果已编制系统参加了ICPR2012的HEp-2荧光型别分类竞赛，取得了优异成绩。本项目有广阔的应用前景，对提高我国的医学水平、保障人民群众的健康安全，有重要的现实意义。

本项目对以HEp-2细胞间接免疫图像为代表的临床医学细胞图像的分割、特征提取与选择、分类识别等关键技术进行了深入研究，为构建细胞图像计算机辅助/自动阅片系统、提升相关疾病临床诊断效率提供了有效的技术支持，具有广阔的应用场景和重要的现实意义。本项目取得了诸多有价值的成果，在细胞图像分割方面：(1)提出一种基于超像素块的Markov随机场细胞图像分割方法。以超像素块作为处理单元，在保持分割精度的前提下，大幅提升分割速度；(2)针对重叠细胞图像分割这个难点问题，提出一种基于Voronoi图的部分重叠细胞图像分割方法。在特征提取与选择方面：针对分割出的细胞区域可提取若干种不同的描述特征，如何将这些不同的特征进行有机选择与组合，构造真正有分辨力的特征集合是本项目的研究关注点之一。结合细胞图像的特点，本项目提出一种改进的RELIEF算法进行细胞图像特征选择。在分类识别方面：(1)提出一种改进的基于判别字典稀疏表示的分类算法，用以高效判别HEp-2细胞图像荧光型别状态；(2)提出一个完整的细胞图像分割、特征提取与选择，分类识别综合处理框架。其他方面：(1)提出一种基于极限分类器ELM的细胞图像样本集不平衡分类问题解决方法；(2)针对临床细胞图像去噪效果判定问题，提出一种基于方法噪声模型的去噪效果评价标准；(3)探索使用诸如GPU、MIC等异构模块加速细胞图像处理的方法。本项目研究期间共发表论文10篇，其中SCI源刊5篇，EI检索4篇。超额完成了任务书中对研究结果的预期（国内外期刊及国际会议发表高质量论文4篇以上）。