当前位置:主页 > 科技论文 > 软件论文 >

基于深度信念网络的肺结节分类研究

发布时间:2019-04-29 18:46
【摘要】:随着医学水平的不断提高,最为普遍的肺部成像技术——CT技术也愈发先进,导致每位就诊人员的肺部CT数据成倍增加,加之患者数量大幅上升,产生海量的CT图像。数据的爆炸式增多是导致对肺癌诊断的漏检、误诊率居高不下的主要原因。目前,由于CAD系统的使用,在一定程度上减少了泛诊,节约了医生的时间与精力,成为医生必不可少的“助手”。针对CAD系统诊断流程的繁琐性,以及最终对结节分类的不准确性,本文主要进行了以下两方面的研究:1.由于结节外形的多样性以及特征的复杂性,在一定程度上会导致结节的过分割,过分割会使有效信息丢失,直接影响诊断的准确度。但如果将原始CT图像(512*512大小)作为任何学习网络的输入,其学习过程的复杂性是不可想象的,甚至是无法完成的。在本文中,首次将目标追踪应用于肺部图像。本文提出的基于超像素的追踪算法是在粒子滤波框架下进行的,首先构建一个基于超像素的肺实质外观模板,然后建立待追踪图像的置信图并设置自适应大小的追踪窗体。在追踪的过程中实时更新模板以保证模板的准确性。保留序列图像中每张CT最优状态时的追踪信息。对原始CT图像进行追踪,快速准确地定位了肺部感兴趣区域肺实质,有效节约了时间成本,为后期的分类做了必要准备。追踪算法可以快速定位出肺部感兴趣区域,有效削弱了CT图像中除肺实质外多余信息的干扰,降低了深度学习应用于肺部疾病诊断的复杂性。2.传统分类方法BP神经网络、支持向量机(SVM)、自生成神经网络(SGNN)等,需人工提取特征,由于不同人有不同的主观标准,因此提取的特征集相差很大。且分类器结构简单,无法运用于大样本数据集,针对上述问题,本文将深度信念网络引入对结节的良恶性诊断中。由于深度信念网络拥有多层非线性结构,对复杂的数据关系有极强的非线性映射能力,且学习过程是由有监督学习与无监督学习交替完成,能更好的完成特征学习与分类任务。在本文中,运用上述追踪方法,得到肺部感兴趣区域,将所有的感兴趣区域图像进行双线性插值,统一图像尺寸,形成深度学习网络的输入数据。接着根据数据集,自定义5层的深度信念网络。对数据进行训练与测试,在训练过程中对网络涉及到的重要参数隐藏层节点数、批数据大小以及样本循环次数等参数进行分析和调优,更有效的实现对结节的分类。
[Abstract]:With the continuous improvement of medical level, the most common lung imaging technology-CT technology is also more advanced, resulting in a doubling of the lung CT data of each patient, coupled with a large increase in the number of patients, resulting in a large number of CT images. The explosion of data is the main reason leading to the missed diagnosis of lung cancer and the high misdiagnosis rate. At present, because of the use of CAD system, the extensive diagnosis has been reduced to a certain extent, the time and energy of doctors have been saved, and it has become an indispensable "assistant" of doctors. In view of the tedious diagnostic process of CAD system and the inaccuracy of classification of nodules in the end, the following two aspects are studied in this paper: 1. Because of the variety of the appearance and the complexity of the features of the nodules, the excessive segmentation of the nodules will lead to the loss of effective information, which directly affects the accuracy of the diagnosis. However, if the original CT image (512? 512) is used as the input of any learning network, the complexity of the learning process is unimaginable or even impossible to complete. In this paper, target tracking is applied to lung image for the first time. The hyperpixel-based tracking algorithm proposed in this paper is carried out in the framework of particle filtering. Firstly, a hyperpixel-based pulmonary parenchyma appearance template is constructed, and then the confidence map of the image to be tracked is established and the tracking form with adaptive size is set up. Update the template in real time during the tracking process to ensure the accuracy of the template. Preserves tracking information for each CT optimal state in a sequential image. After tracing the original CT images, the lung parenchyma of the region of interest was located quickly and accurately, which effectively saved the time cost and made the necessary preparation for the later classification. The tracking algorithm can quickly locate the region of interest of the lungs, effectively weaken the interference of excess information in CT images except pulmonary parenchyma, and reduce the complexity of deep learning applied to the diagnosis of pulmonary diseases. 2. Traditional classification methods, such as BP neural network, support vector machine (SVM), self-generating neural network (SGNN) and so on, need to extract features manually. Because different people have different subjective criteria, the extracted feature sets vary greatly. The classifier is simple in structure and can not be applied to large sample data sets. In order to solve these problems, this paper introduces the depth belief network into the diagnosis of benign and malignant nodules. Because the deep belief network has multi-layer nonlinear structure and strong nonlinear mapping ability to complex data relations, and the learning process is completed alternately by supervised learning and unsupervised learning, the task of feature learning and classification can be completed better. In this paper, the above tracking method is used to obtain the region of interest of the lung. All the images of the region of interest are interpolated bilinear to unify the image size and form the input data of the depth learning network. Then, according to the data set, a 5-layer depth belief network is defined. The data are trained and tested. In the training process, the number of hidden layer nodes, batch data size and the number of cycles of samples are analyzed and optimized to realize the classification of nodules more effectively. 3) the number of hidden layer nodes, the size of batch data and the number of cycles of samples are analyzed and optimized in the process of training and testing.
【学位授予单位】:太原理工大学
【学位级别】:硕士
【学位授予年份】:2017
【分类号】:R734.2;TP391.41

【参考文献】

相关期刊论文 前4条

1 张洋;;基于双线性插值法的图像缩放算法的设计与实现[J];电子设计工程;2016年03期

2 张娟;毛晓波;陈铁军;;运动目标跟踪算法研究综述[J];计算机应用研究;2009年12期

3 顾晓晖;马晓宇;陈卉;;LIDC中肺结节注释信息的提取及数据库的建立[J];数理医药学杂志;2009年02期

4 吕俊,张兴华;几种快速BP算法的比较研究[J];现代电子技术;2003年24期



本文编号:2468446

资料下载
论文发表

本文链接:https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2468446.html


Copyright(c)文论论文网All Rights Reserved | 网站地图 |

版权申明:资料由用户8b145***提供,本站仅收录摘要或目录,作者需要删除请E-mail邮箱bigeng88@qq.com