基于图的医学图像分割方法研究

发布时间：2018-04-22 06:25

本文选题：图像分割 + 随机游走　；参考：《南方医科大学》2017年博士论文

【摘要】：近十多年来基于图的交互式医学图像分割方法受到研究者越来越多的关注,该类方法在临床应用中表现出以下几种优势:一是算法为用户提供了直观而且便捷的交互界面,在足够的交互下可以得到任意的分割结果;二是无需训练步骤或者先验信息,针对各类图像中的不同目标时体现出鲁棒性;三是在自动化种子点检测方法的辅助下,交互式算法可以转化为全自动分割方法。在前人研究的基础上,本文做出的创新与改进工作主要包括两部分:首先提出了一种新型基于随机游走模型的图像分割方法,并将其应用于多模态核磁共振图像中脑肿瘤的分割;其次提出了一种基于半监督谱聚类的迭代式图像分割方法,并将其应用于数字乳腺断层摄影图像中肿瘤病灶的分割。脑肿瘤是一种比较常见且致死率较高的恶性肿瘤,核磁共振成像是针对该类肿瘤最常用的诊断技术。脑肿瘤的多模态核磁共振图像分割是医学图像领域难点和热点问题之一。本文针对多模态脑肿瘤核磁共振图像的分割问题,提出了一种新的基于随机游走模型的分割算法,将随机游走模型拓展到特征空间,根据已标记种子点来计算特征空间中未标记点属于前景或者背景的概率值,并将其做为下一步分割中使用到的预设标签集。并且,为了使算法对初始化具有更高的容错性,作者将传统算法中能量函数的硬约束优化问题转化为非约束优化。作者还提出了一种基于不对称性的种子点生成方法,通过对颅内病灶与肿瘤进行粗略的定位,将上述交互式方法转化为全自动式分割。对比实验显示,当种子点数量下降、种子点位置偏移和错误种子点比例增加时,本文提出算法的性能均优于图割算法和随机游走算法。本研究使用30例多模态神经胶质瘤核磁共振图像对全自动脑肿瘤分割方法进行验证,在分割整体病灶时平均Dice系数为0.852 ± 0.107,在分割肿瘤时平均Dice系数为0.729 ± 0.252。为了有效解决算法中涉及的大规模稀疏线性方程组问题,本研究采用基于GPU技术的CULA线性代数运算工具包作为加速工具,可实现在五分钟之内完成一例分辨率为256× 256 × 181的脑肿瘤图像分割。基于乳腺影像的计算机辅助诊断可帮助医生对图像中可疑肿块和病灶进行检测和分析,是医学图像领域重要课题,而乳腺肿瘤的图像分割是该课题中的关键技术。针对乳腺病灶图像分割问题,本文提出了一种半自动基于半监督谱聚类的迭代式图像分割算法。该算法借鉴了经典算法GrabCut的交互界面,允许用户对像素做不完全标记,只需在目标外输入一个矩形框即可完成算法初始化。该算法提出一种新的相似性图构造方式,将半监督谱聚类在特征空间中的聚类功能和随机游走算法在图像空间中的分割功能整合在同一框架之内,通过迭代式地优化能量函数实现分割过程。本研究使用30例乳腺肿瘤病人(包括15例良性和15例恶性)共计60幅(每例病人包含两个体位)数字乳腺断层图像对上述算法进行验证,实验结果表明该方法可以实现对乳腺肿瘤的分割,分割结果与金标准之间的平均Dice系数为0.798 ± 0.090。本文还对算法中关键参数以及初始化对分割结果的影响做了探讨与分析。
[Abstract]:In recent more than 10 years, the interactive medical image segmentation method based on graph has attracted more and more attention. This method shows the following advantages in clinical application: first, the algorithm provides the user with an intuitive and convenient interface, and can get arbitrary segmentation results under sufficient interaction; two, no training steps are needed. Or prior information is robust to different targets in all kinds of images; three is an automatic seed point detection method, which can be transformed into a full automatic segmentation method. On the basis of previous research, the innovation and improvement of this paper mainly include two parts: first, a new type of base is proposed. The image segmentation method of random walk model is applied to the segmentation of brain tumors in multimodal NMR images. Secondly, an iterative image segmentation method based on semi supervised spectral clustering is proposed and applied to the segmentation of tumor lesions in digital mammography images. Brain tumor is a common and lethal type of tumor. High rate malignant tumor, MRI is the most common diagnostic technique for this kind of tumor. Multi mode NMR image segmentation of brain tumor is one of the difficult and hot issues in the field of medical image. In this paper, a new model based on random walk model is proposed for the segmentation of NMR images of multimodal brain tumors. The algorithm extends the random walk model to the feature space, and calculates the probability values of the foreground or the background of the unmarked points in the feature space according to the marked seed points, and makes it the set of preset tags used in the next step. In order to make the algorithm have higher fault tolerance for the initialization, the author will be able to make the traditional algorithm in the traditional algorithm. The problem of the hard constrained optimization of the quantity function is transformed into a non constrained optimization. The author also proposes a method for the generation of seed points based on the asymmetry. By rough positioning of the intracranial focus and tumor, the interactive method is transformed into an automatic segmentation. The comparison experiment shows that the seed point position is offset and the number of seed points is reduced. When the proportion of wrong seed points is increased, the performance of the algorithm is better than that of the graph cut algorithm and random walk algorithm. This study uses 30 cases of multimodal glioma MRI to verify the full automatic brain tumor segmentation method. The average Dice coefficient is 0.852 + 0.107 when segmenting the whole focus, and the average Dice system in the segmentation of the tumor. The number is 0.729 + 0.252. in order to effectively solve the problem of large-scale sparse linear equations involved in the algorithm. In this study, the CULA linear algebra operation toolkit based on GPU technology is used as an acceleration tool. A case of brain tumor image segmentation with a resolution of 256 * 256 * 181 can be completed within five minutes. Auxiliary diagnosis can help doctors detect and analyze suspicious masses and lesions in images. It is an important subject in the field of medical image, and image segmentation of breast tumor is the key technology in this subject. In this paper, an iterative image segmentation algorithm based on semi supervised spectral clustering is proposed in this paper for image segmentation of breast lesions. The algorithm uses the interactive interface of the classical algorithm GrabCut, allowing the user to make incomplete markings of the pixels. The algorithm initializes the algorithm only by entering a rectangle outside the target. The algorithm proposes a new pattern of similarity graph construction, and the clustering function and random walk algorithm of semi supervised spectrum clustering in the feature space are empty in the image space. In this study, 30 cases of breast cancer (including 15 benign and 15 malignant) were used to verify the above algorithm, and the experimental results showed that the proposed method was used in 30 cases of breast tumor patients (including 15 benign and 15 malignant). The segmentation of breast tumor can be realized. The average Dice coefficient between the segmentation results and the gold standard is 0.798 + 0.090.. This paper also discusses and analyzes the key parameters in the algorithm and the influence of the initialization on the segmentation results.

【学位授予单位】：南方医科大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】