模式识别MCs detection.ppt_第1页
模式识别MCs detection.ppt_第2页
模式识别MCs detection.ppt_第3页
模式识别MCs detection.ppt_第4页
模式识别MCs detection.ppt_第5页
已阅读5页,还剩31页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

1、,Presentation, 2012 BJTU,BEIJING JIAOTONG UNIVERSITY,乳腺X线图像中微钙化点簇的检测 Detection of Clustered Micro-calcifications in Mammograms,姚畅 北京交通大学 2012-3-17,2,. 2 .,内容提要:,一、相关向量机(Relevance Vector Machine, RVM)理论,二、基于自适应核学习的乳腺微钙化点簇检测,3,乳腺X线图像中微钙化点簇的检测 Detection of Clustered Microcalcifications in Mammograms,.

2、3 .,Relevance Vector Machine (RVM):,与support vector machine (SVM) 的关系:,- 都是通过对训练样本集进行训练,获得训练好的模型,实现对新数据的预测。,SVM: - The number of support vectors grows linearly with the size of the training set.,- Predictions are not probabilistic.,- Necessary to estimate the error/margin trade-off parameter C,- The

3、 kernel function K(x; xi ) must satisfy Mercers condition. That is, it must be the continuous symmetric kernel of a positive integral operator.,4,. 4 .,Relevance Vector Machine (RVM):,给定数据集xi, ti,从而,对于待预测数据x*,Relevance Vector Machine (RVM):,为避免过拟合(over-fitting),对w加上先决条件:,从而,,其中,,剩下的工作为求:,Relevance V

4、ector Machine (RVM):,?,(Marginal likelihood),taking logs,get log marginal likelihood:,获得,和,的更新值后,将其返回计算m和,,直到满足收敛标准。,而,RVM for classification:,将分类问题映射为回归问题求解,Development of RVM:,- RVM with adaptive kernel learning (ARVM),In ARVM, learning this model not only computing the posterior distribution of t

5、he weights and estimating the weight precisions , but also estimating the basis function parameters.,ARVM algorithm,Simulation result,Simulation result,Simulation result,Detection of clustered microcalcifications,单个的微钙化点对于乳腺癌的检测没有太大的意义,聚合成簇的微小钙化点簇才是乳腺癌的重要征兆。,17,. 17 .,研究背景:,1 乳腺癌是女性最常见的一种恶性肿瘤,发病率与日俱

6、增, WHO统计,全球每年有超过50万妇女死于乳腺癌,超过120万妇女被诊断患有乳腺癌,我国己成为乳腺癌发病率增长最快的国家之一,以每年3%的速率递增,城市人口乳腺癌发病率己经接近欧洲中等发达国家水平,以上海为例:,1972年,1992年,2000年,2008年,发病率,17人/10万,34人/10万,56.2人/10万,62.5人/10万,中国女性的乳腺癌发病年龄要比西方女性小10岁左右,2 乳腺癌的主要表现形式为微钙化点(Microcalcification, MC)和肿块(Mass),18,. 18 .,研究背景:,3 早发现、早诊断可以很大程度地降低发病率和死亡率, 可以提高病患治疗后

7、的生存期和生活质量, 周期普查,(1次/年,40周岁以上女性),早预防、早发现、早治疗, 乳腺X线摄影术是目前乳腺癌早期诊断中最有效的方法,4 现有技术存在诸多缺点,人工方法存在易疲劳、耗时多、误诊和漏诊率大等问题,乳腺癌计算机辅助检测和诊断(Computer Aided Detection/Diagnosis, CAD)系统能提供有价值的“第二诊断意见”,但存在高假阳性率和局限性,19,. 19 .,MCs研究现状:,MCs检测方法:,基于图像增强方法,随机建模方法,多尺度分解方法,机器学习方法,依据:MCs的灰度比周围像素要亮,差分方法,MCs和周围像素的统计学差异建立检测模型,基于马尔可

8、夫随机域的检测方法,MCs及其周边背景间的频谱差异,小波方法,将MCs检测问题当作一个二分类过程,ANN、SVM、RVM方法,不同方法检测MCs的FROC比较,我国:, 获得了一些成果, 差距明显,任务艰巨,20,. 20 .,面临的问题:,1 现有检测方法大部分都是基于西方女性乳腺X线图像进行的研究,因而对于中国女性常见的致密型乳腺X线图像中可疑病灶的检测,已有方法并不适用,2 理论方法应用于临床CAD系统的可行性、实时性问题,3 提高真阳性率的同时如何解决假阳性率高的问题,4 我国在该领域取得的研究成果还有限,缺乏应用于研究的临床数据库,21,. 21 .,基于自适应核学习的微钙化点簇检测

9、方法,Adaptive kernel learning for detection of clustered microcalcifications,将微钙化点的检测问题转化为一个二分类问题,采用aRVM进行检测,方法流程,预处理:图像来源,1 DDSM: Digital Database for Screening Mammography,the Massachusetts General Hospital (D. Kopans, R. Moore), the University of South Florida (K. Bowyer), and Sandia National Labor

10、atories (P. Kegelmeyer) 2500 cases,2 MIAS: Mammographic Image Analysis Society Digital Database,322 mammograms, each image is 1024 pixels by 1024 pixels,3 A data set collected by the Department of Radiology at the University of Chicago was used to evaluate our proposed method. 141 mammograms from 66

11、 clinical cases, spatial resolution : 0.05 mm/pixel 12-bit grayscale dimension :30005000 pixels.,Preprocessing before extraction the feature,First, in order to suppress the background and thereby restrict the intra-class variations among the training samples, we applied a high-pass filter, which was

12、 designed to be a finite impulse response filter with cutoff frequency and length 41.,and length 41.,Second, normalization,Preprocessing before extraction the feature,apply a template matching procedure with a search window of,pixels to locate the centers of all the manually identified MCs,Feature e

13、xtraction,Due to the fact that individual MCs are small and localized in a mammogram image, we form the input feature by extracting a small window of pixels centered,at the location of interest in an image. Such a choice is to accommodate most of the MCs in size (at resolution of 0.05 mm/pixel) and

14、avoid any potential interference from neighboring MCs.,Kernel functions,A compelling feature of aRVM is that it can automatically optimize the parameters of the kernel functions during training. In this study we consider the following two types of kernel functions:,1. Polynomial kernel,2. Gaussian r

15、adial basis function (RBF) kernel,Note that in RVM the kernel center parameter is placed at the training samples,and those samples with nonzero weight are called relevance vectors.,In aRVM, the center parameters are treated as optimization variables.,Model sparsity,The sparsity prior in is used to c

16、ontrol the complexity of the model. In this study, we use the following different values for the sparsity parameter c, which correspond to several well known model selection criteria:,Determination of the kernel function and the sparsity parameter c,We apply an 5-fold cross validation to determine t

17、he type of the kernel function and the sparsity parameter value.,1 Randomly divide all training samples into 5 equal-sized subsets,2 For each parameter setting, train the classifier model 5 times - during each time, one of the 5 subsets is held out in turn while the remaining 4 subsets are used to t

18、rain the classifier - the trained classifier is then used to classify the held-out subset - In the end, average the classification results to obtain an estimate of the generalization error of the model.,aRVM training,66 cases (141 images),33 cases 79 images,33 cases for testing 62 images,randomly,Tr

19、aining samples,1,578 “MC present” samples,3,422 “MC absent” samples,Relevance vectors (RVs),aRVM RVs,RVM RVs,Test and clusters identification,the trained classifier model is applied to classify each pixel location as “MC present” and “MC absent”; 2) next, the detected “MC present” pixels are grouped

20、 into potential MC objects by a morphological processing procedure, during which isolated spurious pixels are removed; 3) the detected MC objects are grouped into clusters using a clustering criterion.,A true positive (TP) cluster : the objects are connected with nearest-neighbor distances (Dnn) less than 0.4 cm; 2) at least three true MCs should be detecte

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

最新文档

评论

0/150

提交评论