【作者】 赵飞；

【导师】 刘杰；

【作者基本信息】 北京交通大学 ， 生物医学工程， 2017， 硕士

【摘要】 心血管疾病目前是导致人类非意外死亡的主要原因之一,近年来呈现逐渐年轻化的趋势,时刻威胁着人类健康生活。心脏是心血管系统的核心器官,获取心脏的生理病理信息是诊断和治疗心血管疾病的关键。CT成像技术具有成像速度快、成像清晰等优点,是心脏检查的常用手段,CT图像的心脏分割对于心血管疾病的诊疗和定量分析具有重要意义。然而,心脏是实体器官,普通的断层图像序列难以完整地呈现出心脏的生理信息,计算机辅助诊断、介入治疗指导、心脏手术导航等技术的实现通常需要得到完整的心脏解剖结构。医学图像的三维可视化是了解心脏解剖结构的主要手段,而图像分割是实现人体器官三维可视化的数据基础。针对心脏图像分割,国内外研究人员提出了多种分割方法,但是大部分算法是针对心房和心室的分割,并不能满足全心脏分割的要求。同时,心脏的搏动和血液流动过程中得到的CT图像容易出现伪影、边界弱化等问题,影响心脏壁的分割效果,传统手动分割方法需要医务人员的大量经验,分割效率低。因此全心脏的自动分割一直是医学图像处理研究领域具有挑战性的热点问题。本文以医学图像的分割和识别为基础,针对切片中心脏图像与其它组织区别明显,以及心脏CT图像序列中相邻切片间有很高相似性的特点,提出了基于神经网络和图像显著性的全自动分割算法,并对分割后的心脏图像进行了三维可视化重建。本文的主要内容及创新点:1.利用视觉显著性技术,实现基于CT图像的心脏分割。采用了一种图像显著性检测算法,对CT图像进行显著性计算,最后利用显著图像对原CT图像进行分割,得到完整的心脏图像。2.将分割任务分解为定位和分割两部分,采用卷积神经网络和堆叠降噪自编码网络分别进行实现。构建卷积神经网络,实现其对图像中心脏的定位功能,利用定位结果对原心脏CT图像进行裁剪,去掉部分非目标区域;构造了一个堆叠降噪自编码网络,利用手动分割图像对网络进行训练,实现网络对心脏CT图像中属于心脏组织的像素点的分类识别,最后基于分类结果实现对心脏图像的分割。3.对上述分割算法分割后的结果,分别与人工分割结果进行定量评价分析,并采用面绘制和体绘制两种方法对分割结果进行可视化重建。更多还原

【Abstract】 Currently, the cardiovascular disease is one of the main causes of human non-accidental death, it has the high incidence, and no fixed incidence of the law. Now,the cardiovascular disease gradually spread to young people, it threaten the healthy life of mankind always. The heart is the central organ of the cardiovascular system, and obtaining the physiological information of the heart is the key for the diagnosis and treatment of cardiovascular disease. CT imaging technology has the advantages of fast imaging speed, clear imaging and so on. It is the common means of cardiac examination.The segmentation of CT image is important for the diagnosis and treatment of cardiovascular disease. However, the heart is a solid organ, the common fault image sequence is difficult to completely show the physiological information of the heart, and for computer-aided diagnosis, interventional therapy guidance, cardiac surgery and other technologies usually need to get a complete cardiac anatomy. Three-dimensional visualization of medical images is the main means to find out the anatomy of the heart,while image segmentation is the data base to realize the visualization of human organs.For the segmentation of cardiac images, researchers at home and abroad have proposed a variety of segmentation methods,but most of the algorithms focus on the segmentation of the atria and ventricles, and difficult to meet the requirements of the whole heart segmentation. At the same time, due to the heart of the pulse and blood flow process is easily produce artifacts, border weakening and other issues to CT images, affecting the heart wall segmentation effect. Besides, the traditional manual segmentation method requires a lot of experience of medical staff, and the efficiency of segmentation is low. Therefore, the automatic cardiac segmentation has been a challenging hot issue in the field of medical image processing.Based on the segmentation and recognition of medical images, the thesis presents a new method based on neural network and image prominence, which is based on the difference between the cardiac images and other tissues in the slices is obvious, and there is a high similarity between adjacent slices in the cardiac CT image sequence.Then, reconstructed the cardiac images with 3D visualization.The main contents and innovations of this thesis are as follows:1. Using visual significance techniques to achieve cardiac segmentation based on CT images. Taking an image significance detection algorithm to calculate the CT image,and the original CT image is segmented by the significant image to obtain the complete cardiac image.2. The segmentation task is decomposed into two parts: location and segmentation,and using convolution neural network and stack noise reduction self-coding network to achieve the localization and segmentation. Constructed the convolution neural network based on the excellent performance in image classification and target recognition,achieved the positioning function of heart in the image. Then, the image of the original heart CT is cut and the partial non-target area is removed by the positioning result.Besides constructed a stacked denoising autoencoder, and trained it by manually dividing the image to achieve the classification and recognition of the pixels belonging to the heart tissue in the CT image of the heart. Finally, the segmentation of the cardiac image is completed based on the classification result.3. Compared the results of this segmentation algorithm with the artificial segmentation results, and got the quantitative evaluation of these two results. Finally,visualized the segmentation results with two kinds of surface rendering and volume rendering visualization algorithms.更多还原