???jsp.layout.header-default.alt??? ???jsp.layout.header-default.about???
 

Institutional Repository at Tsinghua University >
本科生优秀毕业论文 >
本科生优秀毕业论文(2015) >

???jsp.display-item.identifier??? http://hdl.handle.net/123456789/3670

???metadata.dc.title???: 基于计算机视觉的脑卒中罪犯斑块识别
???metadata.dc.title.alternative???: Atherosclerotic Plaque Magnetic Resonance Imaging Automatic Identification
???metadata.dc.contributor.*???: 张宇兴
交叉信息院
???metadata.dc.subject???: 医疗影像自动判读;机器学习;图像处理;特征提取
medical image automatic examining; machine learning; image processing; feature extraction
???metadata.dc.date.issued???: 18-Mar-2016
???metadata.dc.description.abstract???: 动脉粥样硬化是一种常见的疾病,疾病产生的原因是胆固醇等物质在颈动脉中堆积形成斑块从而阻塞血管。而当斑块从血管内壁上脱落时则有可能引发脑卒中,甚至导致死亡。目前医学上常用磁共振成像的方法来诊断动脉粥样硬化,即通过磁共振扫描获得颈部横截面的图像,再通过医学专家判读来识别颈动脉中是否有斑块然而,由于医院规模的扩大,病人的增加,数据的产生速度终会超过专家的判读速度。因此,我们希望用计算机来替代专家进行对扫描图像的判读。 本文提出了一种用机器学习的算法对颈部磁共振成像图片进行斑块识别的方法,具体而言,本文采用图像处理算法,对输入的磁共振成像图片进行特征提取,然后将每一张图片提取的特征放在一起作为一个数据集,用支持向量机以及不同的核函数进行训练和分类。我们先后尝试了各种不同的特征,包括一个简单的计算血管轴向的包含主要亮度的区间还有比较复杂的通过混合高斯模型拟合出来的亮度分布及其变种,还有灰度直方图等等。除了使用支持向量机,我们还尝试了用深度学习的方法来对图像进行分类,包括正常的神网络,卷积神经网络还有深度信念网络。 最终,通过将不同层的图像进行叠加,我们达到了83%的整体正确率,并且在质量更好的数据集上达到了94%的正确率。最后,本文指出计算机可以帮助医学专家进行磁共扫描图像的判读,这极大的减了人们的工作量,今后甚至还有可能用计算机判读完全替代工判读。这说明基于机器学习法的计算机自动判读很可能在医学上具有极大的应用价值。
Atherosclerosis is a common disease caused by the accumulation of substances in the carotid arteries. Currently people are using magnetic resonance imaging (MRI) to diagnose it, which requires medical experts to examine the images and find out if there is a plaque. Yet as the amount of data generated will finally become too large for medical experts to handle, we hope to use computers to help experts to examine the images. In this paper, we propose a method to identify the plaques in the MRI images of the necks using machine learning algorithms. More precisely, we use image processing algorithm to handle the input images and extract features from them. Then put features of different images together to form a dataset which will be used to train a SVM later with different kernel functions. We have also tried to use deep learning for the classification of images. Finally, we achieve an overall accuracy of 83% and an accuracy of 94% on a dataset of better quality. We conclude that computers are capable of helping medical experts to examine the images from the MRI scanning, which will significantly reduce the work load of the experts. This means automatic image examining based on machine learning algorithms has a great potential in medical industry.
???metadata.dc.identifier.uri???: http://hdl.handle.net/123456789/3670
???org.dspace.app.webui.jsptag.ItemTag.appears???本科生优秀毕业论文(2015)

???org.dspace.app.webui.jsptag.ItemTag.files???

???org.dspace.app.webui.jsptag.ItemTag.file??? ???org.dspace.app.webui.jsptag.ItemTag.filesize??????org.dspace.app.webui.jsptag.ItemTag.fileformat???
047张宇兴2011012151.pdf9.68 MBAdobe PDF???org.dspace.app.webui.jsptag.ItemTag.view???

???jsp.display-item.copyright???

 

Valid XHTML 1.0! ???jsp.layout.footer-default.text??? - ???jsp.layout.footer-default.feedback???