CASIA OpenIR  > 毕业生  > 博士学位论文
深度学习在医学图像分割中的研究与应用
杨雪松
Subtype博士
Thesis Advisor范勇
2019-05-28
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword医学图像分割,深度学习,卷积网络,度量学习,对抗网络,循环神经网络,协同分割
Abstract

图像分割作为医学图像处理中基础性的计算方法,其目的是从生物医学图像中标记出生物结构或器官的物理存在范围,是理解和分析医学图像,进行病理检测、疾病诊断等医学研究工作的数据基础。近年来,医学图像分割问题的研究取得了长足进展,分割算法的计算精度不断提高,应用场景和适用范围得到扩展和衍生。然而,由于医学图像的特殊性,现有分割方法难以满足现代医学分析任务所要求的普适的、自动化的、快速的、便捷的、高精度的图像分割计算。机器学习方法作为一种基于大量数据的统计性计算方法,为解决医学图像分割任务提供了新思路。其通过对训练图像数据的分析和总结,构建特征向量描述像素灰度值、空间位置或其他特征信息(轮廓、尺寸,形状等)的微小差别,并训练分类器推测原始图像中每一个像素点的类别。该类算法的精度依赖于特征向量、分类模型和原始图像数据的契合性,而由于医学图像的复杂性,高精度的机器学习方法并不容易实现,尤其是一些既定的特征提取方法往往较难适应不同类别的医学图像。深度学习模型是近十年来兴起的一种新颖的机器学习方法。该模型利用多层次叠加、互连的网络结构,多尺度地挖掘图像的纹理和空间分布特性,稳定地增强模型对像素类别的分辨能力。相比较传统的机器学习或统计学习方法,深度学习模型结构化、定制化的特征向量能足够精细地表征各类图像的全局分布或局部类别特点,使其可以精准地区分原图像中从属不同类别像素,而特征向量和分类器相互关联的耦合方式也使分割网络更加鲁棒,且易于延伸并吸收一些有益的先验知识。但由于深度学习模型网络结构复杂、层次较多、层次递进关系缺乏严格的数学证明,其应用往往存在网络结构定义复杂,训练缓慢的问题。本文立足于医学图像分割,试图用深度学习模型从多个角度分析解决现有模型疑难问题,丰富深度学习理论和应用基础。论文的主要工作和创新点归纳如下:
1.提出了一种在深度学习框架下的核磁共振图像像素结构化特征的提取方法:
    深度学习利用多级非线性映射,能够挖掘图像像素隐含的纹理和空间分布特性,增强像素的分类能力。本文利用深度网络,对待分割图像像素进行结构化特征提取,并基于此特征为待分割图像的每一个像素构建一个局部的线性分类器,用以判定该图像像素的类别从属。
2.提出了一种基于目标图像和图谱图像相似度的像素级深度神经网络图像分割方法:
    图谱图像作为一种具有高精度分割结果的参考图像在医学图像分割中被广泛地应用,利用图谱图像能快速地对待分割图像像素的类别进行判定。一般来说,该类分割方法的精度在极大程度上依赖于待分割图像和图谱图像之间相似度计算的准确性。本文利用深度卷积网络构建度量计算模型,准确地衡量待分割图像和图谱图像对应像素的相似度,即二者属于同一类别的概率,并利用该相似度数值,融合图谱图像像素的类别信息,推断待分割图像每一个像素的类别。相比较传统相似度计算模型,深度卷积网络能够多角度、多尺度、多粒度地评价待分割图像和图谱图像间的相似或相异程度,从而能较为精确地推断出待分割图像中像素的类别。   
3.提出了一种图像恢复和图像分割协同学习的多任务图像分割方法:
    图像分割任务往往和其他的图像分析任务相关联,如图像恢复。图像恢复能够改善图像的质量,增强像素的可分辨性,提升图像分割的精度;而图像分割的结果明确相邻像素的同类或异类情况,根据同类像素具有相近的灰度值,异类像素具有差异较大的灰度值有效假设,该类别信息将有利于平滑图像恢复的计算结果。据此,本文设计实现一种图像恢复和图像分割同步计算的协同计算模型,利用两种视觉任务的相互激励机制,实现共同提升计算精度,并节省大量的计算时间。
4.{提出了一种基于深度卷积网络的端对端多粒度分割模型:
    基于图谱图像的分割方法依赖于图谱图像标签信息的准确性以及图谱图像和待分割图像质量的匹配性,而适合的图谱图像的收集和制作需要耗费大量的计算成本。本文利用深度卷积网络,构建从待分割图像直接到分割结果的端到端分割模型。该模型根植于传统U型卷积网络,能够很好地解释和表达图像像素灰度分布特征到对应类别标签的非线性映射关系,实现图像的高精度分割。同时该模型利用深度网络多粒度的、树形结构化表达的特点,通过融合处于不同网络深度的分类层的计算结果,进一步提升分割精度。实验证明该模型在实现高精度分割的前提下,省去了耗时的图谱图像构建过程,使医学图像分割计算变得更为简便易用。    
5.提出了一种基于图像空间连续性的平滑分割方法:
    端到端的直接分割模型能够达到较高的分割精度,但该类模型在分割过程中很少或没有考虑待分割图像中相邻像素之间的相互关系。条件随机场模型是常用的一种后处理方法,它能有效地平滑分割结果,实现精度提升。然而现有的分割模型一般将分割操作和平滑后处理作为两种独立的运算,即后处理过程并不参与到分割模型的反馈训练中,这在一定程度上削弱了两种计算过程的依赖关系。据此,本文设计了两种依赖于图像层次间连续性的平滑分割模型。其一,利用循环卷积网络单元在处理序列数据上考虑前后依存关系的优势,在原有卷积分割网络的基础上增加循环单元描述层次间连续关系。其二,利用生成对抗网络GAN的逻辑理念,将基于像素间相互关系的平滑过程作为对抗网络加入到分割网络的训练过程中。通过实验测试,这两种对平滑分割模型均能有效地提升前文设计的基于卷积网络的端到端图像分割模型的计算精度。

Other Abstract

Medical image segmentation is one of the most important fundamental procedures among kinds of medical image analysis tools, it is able to label out regions of biological tissues or organs from images, and provide reliable evidence for medical images understanding. In recent years, the research of medical image segmentation has made a great progress, the precision is getting higher, application scenarios are continuously extending. However, due to unique characteristics of medical images, many segmentation methods are hard to suit automatic, fast, convenient and high-precision segmental calculations. Machine learning methods offer a novol idea for medical image segmentation. By analyzing mass of training data, such methods could extract featuer vectors to distinguish tiny difference of intensity, spatial position or other features (i.e. shape, size, range) of pixels, and build mathematical models to indicate labels of pixels. Generally, the precision of these models raly on conjunction between features, classification models and original image data. However, due to the complexity of medical image, appropriate features are quite difficult to find. Deep learning method is a novel kind of machine learning method that has been provoked lots of attention during last decade. Based on stacked layers, these models are able to dig textural and spatial information of medical images and enhance discriminative ability of pixels. In contrast with traditional machine learning or statistical leaning methods, the structural and customized feature vectors are capable of expressing both global intensity distribution and local discriminative details subtly, thus make it possible to recognize pixels of different categories precisely. Moreover, the coupling of feature vectors and classifiers also makes the segmentation networks more robust and convenient to extend and absorb other useful prior knowledge. However, as the mathmatic relationships between layers are not clear, it is usually hard to define and train a proper model structure. In this thesis, we try several strategies for medical image segmentation based on deep learning models, that may enrich theory and application knowledge of both deep learning models and segmentation tasks. The main contributions of this thesis are summarized as follows:
1.Propose a deep learning  based pixel structual feature extraction method for magnetic resonance image:
Deep learning can unearth hidden textural and spatial information of images based on nonlinear mappings between stacked layers, which is very useful for pixel classifying. Therefore, this thesis proposes a deep learning model to extract structural features of pixels of target images,and builds local classifier, such as supported vector machine, for each pixel based on its features to indicate whether it belongs to foreground or background.
2.Propose a pixel-level deep neural network image segmentation method based on similarity between target image and atlas images:
Atlas is a kind of reference image with manual label mask, which is wildly adopted for medical image segmentation tasks. By using atlas images, categories of pixels of target image can be inferred conveniently. Usually, the performance of atlas images based segmentation methods is hinged on whether similarity between target images and atlas images can be calculated precisely. Thus, this thesis proposes a deep metric learning model to define these similarities, which is probability of pixels belonging to the same category, and involves them with weighted sum formula to indicate category of pixels of target images. By taking advantages of deep learning, deep metric model is capable of evaluating differences between target images and atlas images in multi views, and make it possible to segregate foreground part from target images with high precision.
3.Propose a joint segmentation model for low quality target images:
Image segmentation is always related with other image analyzing tasks, such as image restoration. In particular, image restoration can improve quality of target images, enhance discriminative ability of pixels and promote accuracy of segmentation; and the results of segmentation task indicate category of pixels and show shape of foreground part, based on assumption that pixels of same category should have similar intensities and pixels of different categories should have dissimilar intensities, segmental labels   will be helpful to smooth results of image recovery step. Therefore, this thesis makes use of mutual encouragement of image segmentation and image restoration to improve performance of both tasks. Experiments illuminate validity of joint segmentation model, it can generate high-precision segmental and recovery results synchronously and save almost half of temporal resource-consuming.
4.Propose a deep learning based end-to-end multi-granularitysegmentation method:
The succeed of atlas images based segmentation methods relies on accuracy of manual label masks of atlas and matching rate between target images and atlas images, but such acquired atlas images always need much computational resource to be prepared. Thus, in order to avoid time-consuming for generating of atlas images, this thesis builds an end to end segmentation model, which can map target images to their corresponding label masks directly. Particularly, this model is constructed following by U-net convolutional neural networks, it can interpret and express the nonlinear relationships between intensities and categories of pixels accurately, that makes proposed models to realize high precision segmentation of image conveniently. Moreover, the model can summarize intermediate classification results based on features of layers at different depth and reform them to earn final label mask, that leads further promotion of segmental accuracy. The experiments show deep learning based end to end segmentation method can achieve high performance in comparison of other state of the art methods, and omit procure of fabrication of atlas images, that make it more expedient to solve medical image segregation problems.
5.Propose a smooth segmentation method based on image spatial continuity:
Deep convolutional neural network based end to end segmentation model can achieve satisfying performance, but it does not consider much about relationships between adjacent pixels during segmentation, thus make the segmental results may not have a glossy surface. Conditional random field model is a usually employed post process for automatic label masks smoothing, by making use of joint probability distribution of adjacent pixels, it can get precision improved. However, post process and main segmentation produce are computed independently. In another word, the smoothing step does not participate training and back propagation of segmentation pipeline, that weakens mutual dependency of two processes. Therefore, this thesis builds two types of smooth models: 1). by taking advantages of recurrent neural network, modify convolution layer of basic segmentation model, thus make it possible to represent continuity between layers; 2). build a parallel network to express joint relationships between adjacent pixels, and use adversarial theory to adjust training process of main segmentation model. Experiments verify this strategy can improve segmental results of end to end segmentation model effectively.

Subject Area计算机科学技术
MOST Discipline Catalogue工学
Pages122
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23991
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
杨雪松. 深度学习在医学图像分割中的研究与应用[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(4347KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[杨雪松]'s Articles
Baidu academic
Similar articles in Baidu academic
[杨雪松]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[杨雪松]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.