【作者】 七十三；

【导师】 哈斯苏荣； 阿木古楞；

【作者基本信息】 内蒙古农业大学 ， 基础兽医学， 2008， 硕士

【摘要】 细胞识别和分类计数的自动化是临床医学的重要研究课题之一,而获得一种有效的图像描述量是实现自动识别的关键问题。本实验首先分别建立了7种白细胞原始图像库和吞噬不同数目酵母菌的小鼠腹腔巨噬细胞原始图像库。前者由134种不同显微特征点的白细胞图像组成,后者由128个吞噬不同数目酵母菌的巨噬细胞显微特征组成。对该两种原始图像库中的每种细胞显微图像分别进行平移、旋转、尺度和灰度改变以及加噪声等6～8种变形处理,获取500个不同变形体图像,建立训练样本集图像库。然后利用变形雅可比(p=4,q=3)-傅里叶矩(Pseudo-Jacobi(p=4,q=3)-Fourier Moments ,PJFM’s)对训练样本集图像库中的图像进行数字化描述,并建立了对应的不变矩数据库。为了证明该不变矩对细胞显微图像的数字化描述性能以及抗畸变抗噪声性能,于N=M=10时对白细胞显微图像进行尺度、灰度、旋转、平移等多畸变归-化,并对部分白细胞显微图像分别于无噪声和加噪声条件下进行重建图像实验。另外,分别从训练样本集原始图像和变形体图像,新制备的涂片中随机选取共90个细胞图像,作为实验样本集,用最小平均距离规则进行初步识别实验。为进一步验证该不变矩的数字化描述性能以及对医学领域中的实际应用,建立小鼠腹腔巨噬细胞显微图像库和对应不变矩数据库,借助计算机用最小平均距离规则对25个典型视野的300个细胞图像进行半自动化分类计数。结果表明,用PJFM’s对7种白细胞显微图像进行图像归—化处理后发现,同种细胞图像不同变形体的PJFM’s的值基本相同,而不同种细胞图像的PJFM’s的值有较明显的差异。加噪声图像的重建实验结果显示,重建图像基本恢复了原始图像的主要信息;而且,对部分白细胞的识别实验结果较为理想,平均识别率达98.3%;小鼠腹腔巨噬细胞显微图像的平均分类计数准确率达97%。因此,该不变矩对细胞显微图像具有较好的数字化描述和抗多畸变、抗噪声性能,可作为细胞自动识别及分类计数的特征选择和提取量,以提高临床检测效率。更多还原

【Abstract】 The cell recognition and automatic classification and counting of leucocytes is one of the most important research topics in clinical medicine, and how to get an effective image descriptor is the crucial problem for realizing the automatic identification.Two original image databases, the database for 7 kinds of leucocytes and the database for peritoneal macrophages of mice, were established in advance in this study. The first one was made up 134 leucocytes with different microscopic morphological characteristics, and the second one was comprised 128 peritoneal macrophages swallowed different number of yeast fungus. The training sample set was established by using 500 different versions of cell images which were artificially treated by gradation, rotation and adding background noise. And then Pseudo-Jacobi(p=4,q=3)-Fourier Moments (PJFM’s) was used to describe different microscopic characteristics of cells in the training sample set, and the Invariant Moments Database for these cells were also established.In order to prove the digitally descriptive, anti-distortion and anti-noise properties of the Moments, the microscopic characteristics of leucocytes were theoretically normalized to multi-distortion of scale, intensity, rotation and shift when N=M=10, and the microscopic characteristic images of cells were also reconstructed in our study with or without background noise. Then, 90 leucocytes in experimental sample set, comprised cells of original images and artificially treated versions and newly prepared smears, were initially identified by using the minimum-mean-distance rule. For further confirm the properties of digitalization and application of the this Moment in the medical field, the invariant moments database for microscopic characteristic images of peritoneal macrophage were set up, and semi-automatic classification and counting for 300 macrophages swallowed different number of yeast fungus from 25 pieces of typical visual field was carried out by using the minimum-mean-distance rule.The experimental results showed that the invariant moment values for different kinds of leucocytes were significant difference and the invariant moment values for different versions of same leucocytes were nearly identical after their microscopic images were normalized by PJFM’s. The reconstructed images of leucocytes with background noise indicated that the main information of original images was recovered by finite number of invariant moment values, and the average recognition rate was reached 98.3%. The accurate average classified counting rate of peritoneal macrophages was 97%. Therefore, PJFM’s possesses strong properties of digitalization, anti-multidistortion and anti-noise for microscopic characteristics of different cells, and can be used for feature selection and extraction of microstructure characteristics in microscopic identification, and classification and counting of cells, and also will make a good contribution to improve the efficiency of clinical testing.更多还原