肾小球基底膜TEM图像分割方法的研究

发布时间：2018-05-15 16:04

本文选题：基底膜分割 + 块匹配　；参考：《南方医科大学》2017年硕士论文

【摘要】：慢性肾脏病已成为威胁全球公共健康的重要疾病,肾穿刺病理活检是诊断慢性肾脏疾病的重要手段,借助透射电镜(transmission electron microscopy,TEM)能观察到肾小球细胞亚显微结构的病理改变,从而做出进一步的病理诊断。研究指出,在肾小球细胞的亚显微结构中,肾小球基底膜(glomerular basement membrane,GBM)的变化与慢性肾脏疾病有密切的关系,如薄基底膜病表现为肾小球基底膜弥漫性变薄。因此在病理诊断过程中,医生常常需要对基底膜进行识别和测量。但是GBM的TEM灰度图像纹理复杂,病变种类繁多,且大部分基底膜与周围组织结构对比度较低,依靠肉眼进行识别与测量不仅困难而且耗时。因此,利用计算机图像处理技术对GBM区域进行分割,能更快速直观地观察基底膜的形态,有利于辅助慢性肾脏病的病理诊断。多年来,图像分割算法的研究一直是图像处理领域的研究热点,国内外提出的分割算法也很多,然而针对肾小球基底膜分割算法的研究,是在20多年前才开始逐渐发展起来的。这主要因为生物图像本身带有的复杂性,样品制备过程的多变性,图像对比度差,结构模糊等特点而大大增加了图像分析的难度系数和复杂性,从而在一定程度上使得肾小球基底膜分割算法的发展受到了限制。现阶段已提出的基底膜分割方法可归纳为半自动分割和全自动分割两大类。这些方法主要基于图像的灰度特征、纹理特征、梯度特征等属性分割图像,在处理灰度均匀、形态变化不大的小段基膜时能得到较好的效果,但是当待分割的基底膜具有与周围组织结构对比度较低、自身形状差异性大等特点时,分割性能不稳定,主要原因是因为这些方法对分割对象形状的描述涉及较少,也不能根据已有的分割信息动态调整分割规则。因此,基底膜分割的自动化程度和精细水平仍然需要进一步提升。针对现有方法存在的问题和基于分割基底膜的需求,本文提出了两种方法来完成基底膜的自动分割。方法一是基于块匹配算法的肾小球基底膜自动分割。图像块匹配算法可以有效搜索图像间的相似图像块,但是由于基底膜对比度低、自身形状差异性大的特点,仅仅从一幅参考图像中搜索查询图像的相似块将难以得到最优的匹配结果。而且,当参考图像数量较多时,逐一将查询图像与参考图像进行块匹配,效率是很低的。因此,本文首先针对肾小球基底膜的特点,,将块匹配算法的搜索范围从一幅参考图像扩展到多幅参考图像,并采用了一种改进的搜索方式提高匹配效率。然后开始搜索最优的图像匹配块,最后提取最优匹配块对应的标记匹配块进行加权,重组为肾小球基底膜的初始分割结果。对于匹配结果出现的假阳性问题,本文采用数学形态学的方法对分割结果进行后处理,得到精度更高的结果。方法二是基于随机森林分类器的肾小球基底膜的自动分割。随机森林算法通过bootstrap抽样技术,产生新的训练样本集合,然后对每个bootstrap样本进行决策树建模,生成由k个决策树组合成的随机森林,最后通过投票的方式,对新数据的分类结果进行预测。研究表明,随机森林算法具有较高的预测准确率,对异常值和噪声具有很好的容忍度。但是由于肾小球TEM图像中,不同图像之间存在灰度差异大的问题,使得采用随机森林分类处理海量数据时,会导致部分像素点的分类混乱,致使基底膜的分割准确率不高。本文在随机森林的基础上,引入多重随机森林的概念,从使用一个随机森林进行分类扩展到使用多个随机森林分类,使得当森林数量足够大时,总有一张或多张训练图像的灰度跟待分割图像的灰度接近,进而克服不同图像之间灰度差异带来的分割准确率不高的问题,提高肾小球基底膜的分割准确率。在采集到的500组肾小球透射电镜图像上进行测试,方法一得到的Jaccard系数最低为83%,最高为95%;方法二得到的最低为84.6%,最高为92%。实验结果表明,本文提出的两种图像自动分割方法,在肾小球基底膜的自动分割上取得了精度较高的分割结果,可以为肾活检病理诊断提供有价值的信息。
[Abstract]:Chronic renal disease has become an important disease that threatens the global public health. Renal biopsy is an important means for the diagnosis of chronic renal diseases. The pathological changes of the submicroscopic structure of glomerular cells can be observed by transmission electron microscopy (TEM), and further pathological diagnosis is made. The changes in the glomerular basement membrane (GBM) in the submicroscopic structure of the glomeruli are closely related to the chronic renal disease, such as the thin basement membrane disease showing the diffuse thinning of the glomerular basement membrane. Therefore, the doctors often need to identify and measure the basement membrane during the pathological diagnosis. But the TEM ash of GBM It is difficult and time-consuming to recognize and measure the GBM region by the computer image processing technology, so it can be more quickly and intuitively observed the morphology of the basement membrane, which is beneficial to auxiliary chronic kidney disease. The research of image segmentation algorithm has been a hot topic in the field of image processing for many years. There are also many segmentation algorithms at home and abroad. However, the research on the segmentation algorithm for glomerular basement membrane has been developed more than 20 years ago. This is mainly due to the complexity of the biological image itself and the sample system. The variability of the preparation process, the poor image contrast and the fuzzy structure greatly increase the difficulty and complexity of the image analysis, so that the development of the glomerular basement membrane segmentation algorithm is limited to a certain extent. The proposed method of basement membrane segmentation can be divided into two main parts: semi-automatic segmentation and full automatic segmentation at present. These methods are mainly based on the image's gray features, texture features, gradient features and other attributes to segment the image. It can get better results in the small segment base membrane when the gray level is uniform and the shape is not changed. However, when the base film to be divided has the characteristics of low contrast to the surrounding structure and large difference in its shape, the segmentation property is divided. The main reason for the instability is that these methods have less description of the shape of the segmented objects, and can not dynamically adjust the segmentation rules according to the existing segmentation information. Therefore, the automation and fine level of the basement membrane segmentation still needs to be further improved. In this paper, two methods are proposed to automatically segment the basement membrane. One is the automatic segmentation of the glomerular basement membrane based on block matching algorithm. The image block matching algorithm can effectively search the similar image blocks between the images. But because of the low contrast of the basement membrane and the large difference of the shape of the image, the image block matching algorithm is only searched from a reference image. It is difficult to get the optimal matching result for the similar block of the cable query image. Moreover, when the number of reference images is large, the efficiency is very low when the query image is matched with the reference image one by one. Therefore, this paper first extends the search range from a reference image to a number of references for the characteristics of the glomerular basement membrane. The image is tested and an improved search method is used to improve the matching efficiency. Then the optimal image matching block is searched. Finally, the marker matching block corresponding to the optimal matching block is weighted to restructure the initial segmentation result of the glomerular basement membrane. The mathematical morphology is used in this paper for the false positive problem of the matching results. Methods the result of the segmentation is processed and the result of higher precision is obtained. Method two is based on the automatic segmentation of the glomerular basement membrane based on the random forest classifier. The random forest algorithm produces a new set of training samples by bootstrap sampling, and then models each bootstrap sample to make the combination of K decision tree. The random forest, finally, predicts the results of the new data by voting. The study shows that the random forest algorithm has a high prediction accuracy and has a good tolerance to the outliers and noise. But because of the large scale difference between the different images in the glomerular TEM image, the random forest is used in the random forest. On the basis of the random forest, this paper introduces the concept of multiple random forests on the basis of random forests, and extends from a random forest to the use of multiple random forest classifications to make the total number of forests large enough when the number of forests is large enough. The gray level of one or more training images is close to the gray level of the image to be divided, and then it overcomes the problem of low segmentation accuracy caused by the difference of gray level between different images, and improves the segmentation accuracy of the glomerular basement membrane. In the 500 groups of collected glomerular transmission electron microscopy images, the method one obtains the lowest Jaccard coefficient. 83%, the highest is 95%, and the minimum of method two is 84.6%. The highest 92%. experiment results show that the two image automatic segmentation methods proposed in this paper have obtained high precision segmentation results in the automatic segmentation of the glomerular basement membrane, which can provide valuable information for pathological diagnosis of renal biopsy.

【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R692;TP391.41

【参考文献】