节点文献

基于内容的医学图像检索中特征提取技术研究

Research on Feature Extraction for Content-Based Medical Image Retrieval

【作者】 杨敏

【导师】 张树生;

【作者基本信息】 西北工业大学 , 生物医学工程, 2007, 硕士

【摘要】 随着医学数字化影像设备在临床工作中日益广泛的应用,大量的医学图像数据随之产生。传统的医学档案管理系统采用简单的基于标注的图像数据库甚至完全人工的方法来管理图像数据,严重影响了图像在诊断过程中作用的发挥。如何有效地组织、管理和检索医学图像成为当前迫切需要解决的问题。基于内容的图像检索技术(CBIR,Content-Based Image Retrieval)是利用图像的视觉特征来进行检索,直接对图像内容进行分析并抽取特征,在临床、教学、科研以及医学图像归档和通信系统(PACS)中都有着重要的作用。 由于医学领域许多图像并不包含彩色或在有限条件下采用,因此,在视觉特征中,相比颜色或灰度特征而言,提取纹理和形状特征对医学图像检索显得较为重要。由此,本文重点研究了医学图像的形状特征提取和纹理特征提取。 对于纹理特征,本文利用空间域统计纹理特征描述符灰度共生矩阵实现了对医学图像纹理特征提取,并且将统计法和结构法有机的结合,利用灰度—基元共生矩阵对其进行算法改进。灰度—基元共生矩阵既考虑了图像像素的分布情况,又考虑了像素点周围邻域结构的分布情况。通过实验,验证了该方法鲁棒性好,对旋转不敏感,查询性能得到了改善。 对于形状特征,本文利用小波对图像边缘检测的有效性和相关边界矩对图像的区域和结构描述的统一性,提出了小波多尺度模极大值和相关矩的形状特征提取方法。首先对灰度图像进行小波模极大值变换,得到多尺度的边界图像,再利用6个相关边界矩提取每一个尺度的边界图像的特征,组成图像的特征向量,用欧几里得距离对归一化的特征矢量进行相似性度量,不同的尺度赋予不同的权值。通过实验,验证了该方法可以以一定的精度很好的描述目标物体的形状特征,抗干扰能力强,通用性较高,有效的解决了因图像的平移、尺度、旋转变换等带来的问题。同时针对由Mallat小波多尺度边缘检测方法得到的图像边缘较粗,细节过多的弊端,本文提出了模糊算法进行改进,剔除了冗余数据,得到精确的目标外形,将检索误差降低到最小程度。 最后,本文建立了一个基于内容的医学图像检索实验性原型系统,以一些医学图像为例对所研究的上述算法进行了验证。

【Abstract】 With the rapid improvement of image-based clinical technique, large amounts of digital images are produced everyday. The management of medical image database and how to use those images in Clinical Diagnose Process (CDP) becomes an especially challenge in medical area. Content-Based Image Retrieval plays an important role in clinic, teaching, research and PACS, etc.Many medical images don’t contain color information or are adopted under the limited conditions. Texture and shape features are more important in the medical image retrieval, compared with the features of color or gray. In the context, we focus on the image’s texture and shape features extraction.For texture feature, we propose an algorithm of Gray-Primitive Co-occurrence Matrix that combines the space-statistical texture algorithm of Gray Level Co-occurrence Matrix with the structure distribution of surrounding pixels. Experiment results verify this method is robust, relatively feeble sensitivity to rotation and better query performance.For texture feature, making use of the validity of wavelet edge detection and relative moment depiction of the region and structure, we propose a shape feature extraction based on wavelet and relative moments. First, it transforms the luminance images with wavelet edge detection to get multi-scale edge images, then calculates the six relative moments of every scale. Similarity is given by the Wikipedia distance between two images’ normalized moment vectors. Experiment results verify this method is application independent and effective to solve the problem brought out by image translation, scaling, rotation. Meanwhile, the paper propose blurry algorithm to improve the defect of thick edge and overabundance details. It eliminates redundancy data, obtains accurate object shape and reduces the query error to the least degree.At last, we develop a simple experimental medical image retrieval system based on example image and certificate the following algorithm.

  • 【分类号】R318
  • 【被引频次】13
  • 【下载频次】497
节点文献中: 

本文链接的文献网络图示:

本文的引文网络