论文总字数：23635字

摘 要

随着社会科技的发展和人工智能设备的发明，特别是医学图像处理技术，在医学诊断方面，如何能够准确的找到病因并且快速通过医疗手段进行治疗，图像配准是经过相关的变换，使两幅图像进行对齐配准。只有经过图像配准，才可以对医学图像进行其他处理。

本文采用卷积神经网络（Convolutional Neural Network）和互信息的配准方法，先对医学图像边缘进行提取，使用CNN有效获取边缘特征图，利用边缘特征图去配准抗噪声性比较强，同时可以提高互信息值的计算速度和配准效果。在此基础上，将CNN边缘提取方法与两种经典的边缘检测算子Sobel和 Prewitt比较，最终两种算子效果相当，在图像信噪比上，Sobel要略高于Prewitt算子，但两者都远小于CNN的信噪比，CNN提取效果最好。另外，本文还对比了三种传统的方法，分别是基于 POWELL 算法、PSO优化互信息和梯度互信息的配准，第一种方法CT图像效果略差，X影像和超声影像都能配准得到不错的效果；第二种方法所需时间最短，第三种方法效果适中，而本文选取的基于CNN和互信息的配准方法，在上述三种传统类型的医学图像中都能得到很好的配准效果，配准速度快且精确度高，相关结果证明了将深度学习方法用于与图像配准是可行的。

深度学习的不断发展和进步已经当今社会的一个趋势，在以后的医学领域中，深度学习方法必将受到更加广泛的应用，未来，随着医学图像处理技术的不断发展，对图像配准的要求更高，但基于现有研究水平，还没有提出一种较优的配准方法，快速性和准确性方面仍旧是一个研究的热门方向，希望在以后的研究中，可以将深度学习与医学图像处理技术完美的结合起来，不仅在医学领域更在其他领域发挥作用。

关键词：卷积神经网络（CNN）；互信息；图像配准；边缘提取；优化算法

Research on medical image registration method based on deep learning

Abstract

The continuous progress and development of science and technology and automation equipment in today's era, especially the medical image processing technology, in the aspect of medical diagnosis, how to find the cause of disease accurately and treat it quickly by medical means? Image registration is to align and register two images through relevant transformation. Only after image registration can medical images be processed. First, we use CNN based registration method to extract the edge of the image, and then CNN gets a better edge image.The edge feature map is used to remove the registration noise, which can improve the calculation speed and registration effect of mutual information value. On this basis, comparing CNN edge extraction method with two classical edge detection operators Sobel and Prewitt, the final two operators have the same effect. In image SNR, Sobel is slightly higher than Prewitt operator, but both of them are far lower than CNN's SNR, and CNN extraction effect is the best. In addition, this paper also compares three traditional methods, which are based on power algorithm, PSO optimization mutual information and gradient mutual information registration. The first method has a slightly poor CT image effect, and both X-image and ultrasound image can achieve good results. The second method is fast and the third method is general. The registration method based on CNN and mutual information has good effect on the three traditional medical images, and the registration speed is fast and the accuracy is high. Through our experimental results, it can be concluded that deep learning can be used for image registration.

In the future, with the continuous development of medical image processing technology, the requirements for image registration are higher. However, based on the existing research level, no better registration method has been proposed, in terms of rapidity and accuracy It is still a hot research direction. We hope that in the future we can develop the deep learning and medical image processing technology very well, and it is better to be applied in other fields.

Keywords: convolutional neural network（CNN）; mutual information; picture superimposition;brim extraction; optimization algorithm

目录

摘 要 I

Abstract II

第 1 章 绪论 1

1.1研究背景和意义 1

1.2国内外研究现状 2

1.3论文内容组织 2

第 2 章 图像配准 4

2.1 图像配准基本理论 4

2.2 空间变换 4

2.2.1 刚性变换 4

2.2.2 仿射变换 5

2.3 灰度插值 5

2.3.1 最近邻插值法 5

2.3.2 双线性插值法 6

2.3.3 PV 插值法 7

2.4 相似性测度 8

2.4.1 互信息测度 8

2.4.2 梯度互信息 9

2.5 优化算法 10

2.5.1 Powell 法原理 10

2.5.2 PSO 粒子群算法 10

2.6本章小结 11

第 3 章 卷积神经网络（CNN） 12

3.1 CNN结构 12

3.2 CNN特点 12

3.3 CNN稳定性分析 14

3.4 基于 CNN 和互信息的医学图像配准 16

3.4.1 互信息理论 16

3.4.2 CNN边缘提取与经典边缘检测算子 17

3.4.3 基于CNN的特征提取过程 19

3.5本章小结 20

第 4 章 图像配准结果对比分析 21

4.1 基于 POWELL 算法的配准 21

4.2 基于 PSO（粒子群算法）优化互信息的配准 22

4.3 基于梯度互信息的配准 24

4.4 基于 CNN 和互信息的配准 25

4.5 结果对比分析 27

4.6 本章小结 29

第 5 章 总结和展望 30

参考文献 31

致谢 33

请支付后下载全文，论文总字数：23635字