卷积神经网络在医学图像处理中的应用研究

发布时间：2018-06-27 05:13

本文选题：卷积神经网络 + 医学图像处理　；参考：《湖北工业大学》2017年硕士论文

【摘要】：卷积神经网络对图像分类问题的处理往往优于其他同类型算法,其中的卷积层和子采样层具有能够提取样本特征的功能,而共享权值的特点又极大减少了网络需要训练的参数。科技不断进步的今天,医疗技术也得到了飞快的发展,从中产生的各种病症检查图片更是数不胜数。医师急需摆脱各种繁重的医学图像筛查工作,且如何从无数的病例图中找出某种疾病的相似特征等;如此种种困难不断激励着研究人员,医学图像的研究也渐渐的成为了热点。本文对卷积神经网络应用于两类医学图像展开了研究,其中一类为能反映身体疾病的眼球血丝图,另一类为含有各种级数的脑胶质瘤核磁共振成像图。全文工作如下:(1)首先介绍了卷积神经网络的发展历程,包括国外与国内对其研究的成果,并且详细的说明了卷积神经网络的结构,算法以及推导,完整的阐述了复杂的图像分类问题中应用卷积神经网络的优越性。(2)在经典LeNet-5卷积神经网络结构上实施改进,设计了具有不同卷积核,不同子采样方式与不同分类器的网络结构,并把此结构用于解决识别眼球血丝病症问题。同时在实验环节对输入层样本尺寸,网络的迭代次数进行了探究,对比了改进结构与LeNet-5在使用同一样本数据集情况下的区别,实验表明改进结构能很好的分类眼球血丝所反映的病症。(3)根据脑胶质瘤多层图片的特点,并基于眼球血丝网络模型,设计出多列卷积神经网络结构:每一层的脑胶质瘤样本作为每一列的输入,同时增加了卷积和子采样层的层数,并使用Maxout激活函数替代了传统神经网络中经常使用的Sigmoid函数。实验部分取多列结构与单列结构,人工提取特征方式实施了对比,结果凸显了多列结构在胶质瘤分级上的优势;此外还对样本进行优化处理,进一步提高了分级准确度。最后本文对多列卷积神经网络计算进行了可视化处理,从视觉方面解释了每层的工作过程。
[Abstract]:Convolution neural network is usually superior to other similar algorithms in image classification problem. The convolution layer and sub-sampling layer can extract the feature of samples, and the characteristics of shared weights greatly reduce the network parameters that need to be trained. With the development of science and technology, medical technology is developing rapidly. Doctors urgently need to get rid of all kinds of heavy medical image screening work, and how to find out the similar characteristics of a disease from countless case maps, and so on; such difficulties continue to inspire researchers, medical image research has gradually become a hot spot. In this paper, the application of convolutional neural networks to two kinds of medical images is studied. One is a hemodigram of the eyeball which can reflect the body disease, the other is a magnetic resonance imaging of glioma with various stages. The main work of this paper is as follows: (1) the development of convolutional neural network is introduced, including the research results both at home and abroad, and the structure, algorithm and derivation of convolutional neural network are explained in detail. The advantages of applying convolution neural network in complex image classification problem are discussed. (2) the network structure with different convolution kernel, different subsampling method and different classifier is designed by improving the classical LeNet-5 convolutional neural network structure. And this structure is used to solve the problem of identifying ocular blood disease. At the same time, the sample size of the input layer and the number of iterations of the network are explored in the experiment, and the difference between the improved structure and LeNet-5 in the case of using the same sample data set is compared. The experimental results show that the improved structure can well classify the diseases reflected by the blood filaments of the eyeball. (3) according to the characteristics of the multilayer images of gliomas and based on the network model of the blood filaments of the eyeball, The multi-column convolution neural network structure is designed: each layer of glioma samples is used as the input of each column and the number of layers of convolution and sub-sampling layers is increased. The Maxout activation function is used to replace the Sigmoid function which is often used in the traditional neural network. In the experiment, the multi-column structure and single-row structure are compared, and the results show the advantages of multi-column structure in glioma classification. In addition, the sample is optimized to further improve the classification accuracy. Finally, the multicolumn convolution neural network computation is visualized, and the working process of each layer is explained from the visual point of view.
【学位授予单位】：湖北工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】