基于Snake的图象分割与癌细胞识别方法研究

【作者】 胡敏；

【导师】 平西建；

【作者基本信息】 解放军信息工程大学 ， 信号与信息处理， 2005， 博士

【摘要】 随着医学图象可视化技术的发展和各种医学成像模式的出现，医学图象自动分析和处理已成为图象工程领域和生物医学工程领域一个重要的研究方向。作为医学图象处理中的一个热点问题，细胞图象的自动分析和识别一直受到人们的普遍重视。由于细胞核浆形态多样，细胞涂片中存在细胞重叠与杂质污染，染色不均匀，涂片细胞图象的高精度分割与恶变性状特征提取成为细胞图象处理和癌细胞定量分析与识别中的难点课题。 本文研究细胞图象分析技术和癌细胞识别方法。针对食管涂片的细胞图象，应用图象分析和模式识别技术研究细胞的分割方法和细胞形态、颜色、纹理特征的提取，以及癌细胞分类识别技术。本文的研究成果主要体现在以下几个方面： 1、提出了一种基于模糊灰度一致性的Snake生长模型。针对传统Snake须将初始轮廓曲线置于真实边界附近的缺点，该模型在能量函数中增加了一项基于像素点与目标灰度一致性模糊度量的生长能量，使得能量优化过程不易受局部极小值的影响，具有较强的抗噪能力。轮廓曲线采用极坐标描述，计算简便。实验结果表明，该模型分割效果良好，分割性能稳定。 2、针对细胞核边界重叠和模糊现象，提出了一种基于信息融合的新的Snake生长模型，并构成了一种彩色细胞核分割方法。该方法充分利用细胞图象的先验信息，对细胞核进行椭圆拟合和边界重叠(污染)信息估计，基于椭圆边界和不同区域的颜色分布特点，建立多个模糊度量函数分别从几何关系和颜色一致性上描述像素点对细胞核的隶属程度，然后融合边界估计信息和各种模糊度量，建立新的Snake生长模型实现细胞核的分割。椭圆信息增强了对重叠或模糊部分的边界跟踪能力，多种信息的融合改善了分割效果。实验结果表明，新方法分割精度进一步提高，分割性能更稳定。 3、提出多种细胞核恶变性状的特征分析方法。针对癌细胞核染色颗粒特征明显的特点，提出一种基于形态学颗粒分析的纹理描述方法(MSGF方法)。该方法构造一种二维粒度分布图对二值图象作颗粒元素分解，以颗粒元素的数量、尺度分布和几何特征参数代替传统SGF纹理分析中的连通区域特征参数，对细胞核染色颗粒特征具有更好的描述能力。此外，本文还采用曲率熵描述细胞核的形状不规则程度，给出了改进的Tamura纹理粗糙度和对比度参数描述细胞核的染色质粗细程度。 4、在应用上述方法对细胞图象进行精确分割和恶变性状分析的基础上，对提取出的一系列细胞核形状、颜色和纹理特征参数，分别应用贝叶斯分类法和k-近邻法进行癌变细胞与非癌变细胞的分类识别实验。实验表明，单细胞的分类正确率达到86％以上，在对少量样本作拒分决策的情况下可以获得更高的分类正确率：与传统的SGF特征和癌细胞识别中使用较多的GLCM特征相比，本文提出的MSGF特征描述细胞核恶变性状更有效，分更多还原

【Abstract】 Medical image processing has become an important area in image analysis and biomedical engineering fields due to the rapid development of medical imaging technology. As a hot issue of medical image processing, cell image analysis has attracted widely the attentions of researchers. The quantitative analysis and malignity detection of cell images obtained from smears is usually taken as a difficult problem, because of the diversity of contained tissues, uneven staining, and overlapped cell clusters in the cell images.This thesis focuses on the quantitative cytological analysis and automated cancerous cell recognition technology, aiming at the cell images obtained from esophageal smears. Based on the in depth research on image analysis, pattern recognition theories, and cytological pathology knowledge, we make a systematical and comprehensive study on the technologies of cell segmentation, cytological malignant feature description and cancerous cell recognition. The main contributions of this thesis are summarized as follows:(1) A growing snake based on fuzzy intensity consistency measure is proposed. To solve the initialization problem of the traditional snake, we improve the energy function by adding an adaptive growing energy term defined by the pixel’s fuzzy intensity consistency measure. The growing snake has strong anti-noise ability and low computation cost. The experiments show that the proposed snake model has encouraging segmentation results and stable performance.(2) Aiming at the overlapped or blurred nucleus edges, we propose a novel information fusion based growing snake to segment color cell nucleus. Utilizing adequately the prior knowledge of cell images, we firstly perform ellipse fitting on the nucleus and give an estimation on the edge superposition status. Based on the detected ellipse and tristimulus distribution characteristics of different regions, we define several fuzzy measurements to describe the degrees of the pixel belonging to the nucleus geometrically and tinctorally. At last, we fuse these fuzzy measurements with different methods and build a new growing snake to segment the nucleus. The ellipse information enhances the boundary tracking ability for the overlapped or blurred edges. The fusion of various information improves the segmentation accuracy and performance stability.(3) Several feature description methods are proposed to analysis the malignant characteristics of nucleus. To effectively analysis the granularity of nucleus chromomere, we proposed a morphological granularity analysis method, called MSGF method. The MSGF method constructs a 2D granularity distribution graph for the bi-level image and performs granularity element decomposition on it. The size distribution and topological feature parameters of the granularity elements are used to replace the connective region parameters in the traditional SGF texture description method. In additional, we use the curvature entropy to measure the更多还原