论文总字数:30887字
摘 要
核磁共振图像(Magnetic Resonance Imaging, MRI)的脑组织提取是将核磁共振脑图像中的脑部组织与非脑组织分离,去除非脑组织,也被称为头骨剥离或全脑分割。脑组织的准确提取是神经图像处理中一个重要的步骤,对于疾病诊断与治疗、认知研究等具有重要意义。对于人脑磁共振图像,手动提取脑组织是一项繁琐耗时的工作,过去的几十年内有许多脑组织提取的方法,这些方法通常需要首先从数据中提取特征,还有很多参数需要设置,因此,这些方法只能称为半自动的方法。
卷积神经网络是深度学习的重要组成部分,可以很好的学习到数据在空间上的特征,在图像和语音识别方面取得了很好的效果。最近流行的全卷积神经网络在图像自动分割研究中也取得了很好的应用效果。在目前常用的全卷积网络模型中,首先通过卷积进行特征提取,然后将全连接层卷积化,上采样部分融合了多层的信息,最后得到与输入图像尺寸相同的分割结果作为输出。然而,这类端到端的网络模型在下采样等网络传递过程中细节特征会丢失,在上采样时这些细节信息难以恢复,会影响脑组织提取的精度。为了应对这一问题,本文将经典的Inception模块引入到了全卷积网络模型中,可以有效地克服细节信息丢失的问题。
为了评估本文提出算法的效果,本文选用了脑组织自动提取中通用的IBSR18与LPBA40数据集进行实现。本文提出的算法与三种方法进行了比较,包括传统全卷积神经网络与两种经典的脑组织自动提取工具。通过定性比较与定量分析(通过计算Dice系数、sensitivity、specificity以及绘制absolute error map作为评价指标),验证了本文提出算法的优越性。
关键词:大脑磁共振图像、脑组织提取、卷积神经网络、语义分割、inception
Abstract
Brain extraction from Magnetic resonance imaging (MRI) is the separation of brain tissue from non-brain tissue in MRI brain images and removal of extra brain tissue, also known as skull stripping or whole brain segmentation. Accurate extraction of brain tissue is an important step in neural image processing. Accurate extraction of brain tissue is of great significance for disease diagnosis and treatment, cognitive research and so on. For human brain magnetic resonance images, manual extraction of brain tissue is a tedious and time-consuming work. In the past few decades, there are many methods of extracting brain tissue. These methods usually need to extract features from the data first, and many parameters need to be set. Therefore, these methods can only be called semi-automatic method.
Convolution neural network is an important part of deep learning. It has a good ability to learn the characteristics of data in space very well. It has achieved good results in image and speech recognition. Recently, the popular fully convolutional networks (FCN) has also achieved good results in automatic image segmentation. In the current full convolution network model, the feature is extracted by convolution first, then the full connect layers are replaced by convolution, the up sampling part fuse multi-layer information. Finally, the segmentation result with the same size of input image is obtained as output. However, this kind of end to end network model lost detail feature in the process of network transmission such as down sampling. It is difficult to recover the details when up sampling, which will affect the accuracy of brain tissue extraction. In order to deal with this problem, this paper introduces the classical inception module to the fully convolutional network model, which can effectively overcome the problem of missing details.
In order to evaluate the effectiveness of the algorithm proposed in this paper, IBSR18 and LPBA40 data sets used in automatic brain tissue extraction are selected. The algorithm proposed in this paper is compared with three methods, including the traditional fully convolutional network and two classical automatic brain tissue extraction tools. Through the qualitative comparison and quantitative analysis (by calculating the Dice coefficient, sensitivity, specificity, and drawing absolute error map as the evaluation index), the superiority of the proposed algorithm is verified.
KEY WORDS: brain MRI, brain tissue extraction, convolutional neural network, Semantic Segmentation, inception
目录
摘要 I
Abstract II
第一章 绪论 1
1.1 研究背景 1
1.1.1 大脑成像技术 1
1.1.2 磁共振图像脑组织提取 2
1.2 研究现状 2
1.3 内容安排 4
第二章 卷积神经网络 5
2.1 卷积神经网络基础 5
2.1.1 卷积层 5
2.1.2 激活函数 7
2.1.3 池化层 7
2.1.4 全连接层 8
2.1.5 Softmax分类器 8
2.2 本章小结 8
第三章 模型与方法 9
3.1 引言 9
3.2 基于全卷积网络实现磁共振图像自动提取脑组织 9
3.2.1 全卷积网络结构 9
3.2.2 全局特征与局部特征结合 11
3.3 在全卷积网络中引入Inception模块 11
3.3.1 Inception模块 12
3.3.2 加入inception模块的网络结构 14
3.4 本章小结 16
第四章 实验及结果 17
4.1 数据集及预处理 17
4.1.1 数据集 17
4.1.2 数据处理 18
4.2 评价指标 19
4.3 实验环境及参数 19
4.4 实验结果 20
4.5 本章小结 23
第五章 总结与展望 24
5.1 研究总结 24
5.2 研究展望 24
致谢 26
参考文献 27
绪论
研究背景
大脑是人类神经系统的中枢器官,是人体最复杂的器官,它控制身体的大部分活动,处理、整合和协调从感觉器官接收的信息,并对发送给身体其余部分的指令作出决定。大脑被头部的头骨所包围并受到保护。由于大脑的重要性,对于大脑健康的关注必不可少。环境变化、人类工作和生活方式的改变甚至交通事故等都会导致脑部疾病的发生,例如脑萎缩、癫痫、颅脑外伤、肿瘤等。因此,对大脑的研究和分析一直是研究的热点。
得益于医学影像技术的发展,可以通过各种医学成像技术获得大脑图像,图像语义分割是医学图像分析中最重要的任务之一,在许多临床应用中往往是第一步也是最关键的一步。在脑磁共振图像分析中,脑组织提取(全脑分割)通常用于测量和可视化大脑的解剖结构,用于分析脑的变化,描绘病理区域,以及用于手术规划和图像引导的干预。
大脑成像技术
医学影像技术通过某种介质与人体的作用,将人体器官以图像的形式体现出来。对于大脑成像,目前常用的成像技术包括以下几个:
剩余内容已隐藏,请支付后下载全文,论文总字数:30887字
该课题毕业论文、开题报告、外文翻译、程序设计、图纸设计等资料可联系客服协助查找;