【作者】 裴政；

【导师】 康雁；

【作者基本信息】 东北大学 ， 生物医学工程， 2009， 硕士

【摘要】 近年来,数字化X线摄影,尤其是直接数字化X线摄影系统的应用,使传统的X线摄影技术进入数字化领域。它与传统的X线摄影相比,具有更高的影像质量,包含更多的影像信息。这种数字化信息经过后处理后,可以获得更多的科研、教育、诊断应用。本文以数据库检索DR胸部图片为目的,研究了关于X光胸片的几个关键图像处理和分析技术。主要包括肺部分割、肋骨分割和基于纹理的数据库检索。具体内容如下：对于DR胸片来说,肺部区域的分割是计算机辅助检测胸腔疾病的首要步骤。本文提出两种自动检测肺部边界点的位置的方法。第一种方法首先在图像中计算一系列的参考线来检测肺部区域在整个图像中的大致位置。然后,在对图像进行平滑处理后,在图像水平、垂直方向上,通过求得所在行或列范围内的候选点的灰度值差的极大、极小值来确定边界点的位置,差值为正的代表亮度由低到高,差值为负的代表亮度由高到低。在肺顶区域去除了锁骨的影响。第二种方法通过使用一系列的感兴趣的矩形区域来确定边界点的位置。最后,为了提高边界点的准确性,肺部近似边界的位置可以通过活动轮廓模型的方法被调整,也可以直接使用插值方法将边界点连起,删除肺野的外部区域。使用上述两种方法,对实际采集的DR图片进行测试,能够较好的分割出肺部边界。在DR数字胸片图像中,肋骨区域对于准确分析肺部区域的纹理存在着较大的影响,在分析肺部区域之前,先将肋骨区域分割出来,然后再对分割后的肺部区域进行纹理分析,有利于DR数字胸片图像的准确分类,也对后继的病理诊断具有非常重要的意义。本文使用了三种分割肋骨的方法,分别是K—聚类均值法,高斯阈值曲面方法和霍夫变换方法,霍夫变换的方法结果好于其他的方法,在使用霍夫变换前需要对图像增强,细化肋骨区域,根据霍夫变换曲线受间断的影响较小,在已知曲线形状的条件下,利用分散的边缘点进行曲线逼近的优点,得到肋骨区域中心线,最后得到肋骨区域。本文最后提出了基于胸部图像纹理特征的检索系统,使用共生矩阵的纹理特征检索方法,在检索前提出了减少肋骨阴影对最后的肺部区域的纹理检测的影响的算法,通过查全率和查准率评价检索结果。更多还原

【Abstract】 Recent years, digital radiography, especially the application of the DR (direct digital radiography) system leads the traditional radiography into digital field. Comparing with traditional radiography image, direct digital radioghy image has higher quantity and includes more information. After image manipulation, the digital information can be applied in more clinical diagnosis.This paper is based on the clinical application of CAD, and study several key technologies of image processing and analysis in X-ray. The main content includs lung segmentation, rib segmentation and texture-based image retrival techniques.For DR chest images, the precise location of the lungs is the chief mission of a computer-aided diagnosis system to detecte chest disease. This paper proposed two methods to automatically detect the location of the lung boundary points. The first method calculates a set of reference lines to determine the relative position of the lungs in the image. Mediastinal and costal edge points are detected on the image which is the results of the application of a 1-D first derivative Gaussian filter to image rows. In this filtered image, positive values are associated with increasing intensity transitions, while negative values correspond to decreasing variations. Therefore, mediastinal and costal edge points are represented by maxima and minima derivative respectively. The methodology to detect top and bottom boundaries is similar to the one described for lateral border points, except for the use of image columns instead of image rows. To discriminate between lung and clavicle edges, all points locating beyond the exterior part of the costal boundary are eliminated. For improving the detection of lung contours, a method based on the use of active contours models was developed. Vertical and horizontal rectangular regions of interest (ROIs) are studied to identify the preliminary edge. These points are approximations to the lung edges which are adjusted using the active contours models. Interpolation method is used to directly connect the boundary points. The use of the above-mentioned two methods can get better lung border segmentation.Rib region influences the texture analysis of lung area. We need to extracte the rib region before analyzing the texture feature in lung region. Several segmentation methods are presented to detect the ribs in digital chest radiographs, including K—means clustering, Gaussian curve plane threshold methods and hough transform. The evaluations of their results are present in the end of the paper. The experimental results indicate that hough transform method is more effective than the others to detect the ribs in digital chest radiographs. The hough transform is carried out after enhancing the image and thinning the image. Hough transform curve has less effect on the intermittent and has an advantage of curve approximation, so it can get the centerline of the ribs region.Finally, analysis of texture feature extracted by gray level co-occurrence matrix is presented. The rib borders are plotted on the original input image. The difference between rib shadow and lung part without rib shadow is subtracted from the rib shadow since the rib shadow has higher. The recall and precision ratio reveals the retrieval result.更多还原