【作者】 陈健美；

【导师】 宋顺林；

【作者基本信息】 江苏大学 ， 计算机应用技术， 2008， 博士

【摘要】 图像识别是根据图像数据特征,利用识别理论与方法对图像进行分类的过程,属于图像处理和模式识别的研究范畴。医学图像识别是医学影像自动诊断的核心内容,是国内外医学领域重点研究的方向。医学图像中蕴含着丰富的人体图像特征信息和规则,其高分辨率、数据的海量性、图像特征表达的复杂性等特点,使得医学图像识别研究面临挑战。研究和探索适合于医学图像自动识别的技术、方法及其算法等医学图像识别的理论和实践问题具有重要而现实的意义,对辅助医生进行医学图像临床诊断具有重大实用价值。论文以医学图像数据为研究对象,在总结国内外关于医学图像识别研究的关键问题、技术方法和研究进展的基础上,针对医学图像数据特征的多维性、复杂性等特点,系统研究了医学图像识别问题,并提出基于密度聚类、多特征融合、融合特征关联的医学图像识别技术、方法及其算法。主要内容体现在以下几个方面。(1)医学图像数据密度分布研究。论文研究了非参数密度函数构造、混合密度函数构造,并将其应用到医学图像识别研究领域。图像象素的灰度及其密度是表达医学图像特征的主要内容,文章深入研究了医学图像的核密度估计函数和混合密度函数,提出适合医学图像数据的核密度估计模型和高斯混合密度估计模型。(2)基于密度聚类的医学图像识别研究。论文从聚类分析的角度出发,根据医学图像的核密度估计模型和高斯混合密度估计模型,提出了基于核密度模型的医学图像爬山聚类算法和基于高斯混合密度模型的医学图像数据加权聚类算法。应用这两种聚类算法进行了实验研究,结果表明,算法能够较好地分类出医学图像中有语义的人体器官,达到了医学图像有效识别的效果。这种基于密度聚类的医学图像识别为医学图像自动分割提供了新的技术支持。(3)基于多特征融合的医学图像识别研究。医学图像的许多特征都能在一定程度上表达医学图像内容,论文围绕多特征问题,试图从特征融合的角度,研究医学图像的识别问题。从表达医学图像的内容特征出发,论文系统研究了图像特征级的数据融合问题,提出医学图像特征数据融合的框架和基于多特征融合的医学图像识别方法和实现技术。(4)基于特征融合的医学图像关联识别研究。论文提出了一种基于特征融合的医学图像关联识别方法,该方法同时考虑医学图像的所有属性,在训练样本上挖掘频繁属性集和类标签之间的强关联规则,关联规则中的频繁项目集挖掘采用频繁闭项目集挖掘方法,并利用这些强关联规则构造分类器,从而可以判断给定医学图像是否正常,因而能进一步提高医生诊断病情的准确性。本文提出的基于密度聚类的医学图像识别方法及其算法、基于多特征融合的医学图像识别方法及其算法、基于融合特征的医学图像关联识别方法及其算法等创新性研究成果,对医学图像识别研究、医学图像自动诊断和临床医学早期诊断都具有重要意义。更多还原

【Abstract】 Image recognition is such a process of classifying images on the features of image by recognition theory and methods,which belongs to the scope of image processing and pattern recognition.Medical image recognition is the vital content of automatic diagnosis by medical images and also an important research direction for medical fields all over the world.A lot of feature information and rules are hiddened in the human medical images.Those images have characteristics of high resolution, huge volumes of data and complicated expression of image features,therefor many challenges must be faced with to recognize them.It is realistically significant to investigate and explore theories and practiccal techniques,methods and algorithms to recognize medical images automatically.This has great practical value for assisting doctors to make clinical diagnosis.Medical images are the objects for studying in this paper On the basis of summary about research works for main hardships,techniques,methods,state of the art and developments of image recognition,this paper presents algorithms,methods and techniques of medical image recognition based on density clustering, multi-feature fuse,fused feature association rule,which can deal with the characteristics of high dimension and complicacy for medical images.The main contents include following several aspects.(1) Medical image data density distribution.Nonparameter density functions and mixture density functions have been constructed to recognize medical images.The gray and its density of each pixel in image are the main expression content.Kernel density estimations and mixture density functions are deeply probed and heir corresponding models are constructed to estimate medical image data.(2) Medical image recognition based on density clustering.From the view of clustering analysis,algorithms of hill-climbing clustering based on kernel density model and weighted clustering based on Gaussian mixture density model for medical image data are presented.Both of the clustering algorithms are applied to our experiment.The results show that they can classify the semantic information of human medical images and have the effects of recognizing medical images.This method of medical image recognition offers new technique supports for medical image automatically segmentation.(3) Medical image recognition based on multi-feature fuse.Research works have been done around the topic of multi-feature fuse for medical images.Many features of medical images can express contents of medical images,so this paper try to recognize medical image from the image features fuse.On the basis of features expression for medical image content,data fuse on the level of image features is studied systematically.Framework of fusing medical image features data is designed. Methods and techniques based on multi-feature fuse are implemented.(4) Medical image association rule recognition based on fused features.The method of medical image association rule recognition based on fused feature is provided.All feature items of medical image are considered in this method.Frequent item sets in training samples are mined by our frequent close items set method and strong association rules between class labels are mined to construct classifiers which can make the conclusion that the medical image is normal or not.This method can help doctors to improve the accuracy of diagnosis.The many innovate research fruits,methods and algorithms for medical image recognition based on density clustering and multi-features fuse,and association rule recognition based on features fused have,offer new theory basis and operation methods for medical image recognition,automatically diagnosis and forepart clinic diagnosis.更多还原