【作者】 刘婷；

【导师】 冯乃章； 孙明健；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2016， 博士

【摘要】 近年来,光声成像因其兼具高灵敏的光学对比度和大超声穿透深度的优点而越来越受到临床与产业的关注,成为当前生物医学领域发展最快的医学成像技术之一。光声显微成像技术以其独特的聚焦成像方式,可以实现单根毛细血管的高分辨率成像,非常适合于微循环系统病变的研究。但是,传统的光声显微成像技术属于扫描成像技术,时间分辨率不高,成像速度慢,难以满足生物学研究中快速实时成像的要求;另外,数据量大,对系统的硬件要求较高,严重阻碍了该技术向临床应用和转化的进程。为了解决传统光声显微成像系统由于点对点的机械扫描方式所导致的成像速度慢、数据量大和不稳定等问题,提出非均匀稀疏采样方案以提高系统的时间分辨率,实现光声数据的快速采样。通过对肿瘤和血管光声显微图像非均匀特性的分析,提出一种更适合的非均匀采样方案,将扫描位置限制于感兴趣区域或成分,进一步减少采样点数,同时又保证一定的光声显微图像信噪比。具体的,针对肿瘤和血管光声显微图像,探讨非均匀采样的必要性,并将图像分为感兴趣区域集中类型和不集中两类。针对感兴趣区域集中类型的图像,提出基于边膨胀图理论的非均匀采样方案,该方案采取重点区域采样的思想;针对感兴趣区域不集中的图像,提出了一种基于低秩矩阵近似的非均匀采样方案,采用的是主成分多采的思想。这两种非均匀采样方案均可以进一步提高光声显微数据的完备性,降低采样点数,以加速采样过程,更好的恢复光声显微图像。为了实现光声显微图像的快速精确恢复,研究图像的恢复模型。首先为了快速获取光声显微图像,根据低秩矩阵近似理论设计基于Go Dec的快速图像恢复算法;并且基于实际的光声实验数据,分析稀疏采样模板与图像恢复算法之间的关系,给出拟合的经验公式,指导图像恢复算法的参数选择。进而针对低秩矩阵恢复问题,采用核范数最小化求解的凸松弛方法在ADMM框架下进行求解,并深入分析低秩约束图像恢复算法存在的边缘保持效果差和恢复图像存在伪影的缺点;最后,为了提高低秩恢复算法的精度,更准确的恢复图像细节信息,引入表征图像结构先验信息的稀疏约束项,构成低秩和稀疏矩阵恢复问题,并在ADMM框架下推导一种通用的适合低秩和稀疏联合求解的多约束矩阵恢复问题的非精确ADMM求解方法,实现光声显微图像的高分辨率恢复。在研究了非均匀稀疏采样方案和图像恢复算法的基础上,搭建快速扫描光声显微成像系统。该系统通过深入分析传统光学分辨率光声显微成像系统,引入二维光扫描仪,实现光扫描模式的光声显微成像,克服机械扫描的不稳定性;另外,在不增加系统成本的基础上,通过光扫描仪的合理控制以及各个试验部分的同步控制,实现光声数据的稀疏采样,即光声数据的快速采集,提高时间分辨率。随后,利用搭建的光声显微成像系统,设计针对肿瘤和血管目标的实际实验,验证稀疏采样方案和图像恢复算法的有效性。将搭建的光声显微成像系统应用于微循环血管疾病的可视化图像显示,为了全面精确量化光声显微血管图像的特征,给出表征血管疾病的量化参数,提出一种基于多尺度Hessian滤波器的光声血管图像量化方法。该方法中,首先针对光声显微血管图像的多尺度特征,提出基于多尺度Hessian滤波器的复合光声图像血管分割方法,获得精确的血管分割结果;在分割图像的基础上,计算全面量化血管位置及形态信息的四个量化参数:血管密度、血管长度分数、血管直径和分形维数;提出一种局部小区域的量化方法,给出量化参数的量化彩图,实现血管信息的局部量化。最后将量化方法与快速扫描光声显微成像系统相结合,应用于血管目标的成像及量化,进一步验证系统的有效性。更多还原

【Abstract】 In recent years,more and more clinical and industrial attentions have been paid to Photoacoustic Imaging(PAI)for its high optical contrast and large ultrasound penetration depth.It has emerged as one of the fastest growing medical imaging technology in biomedical applications.With its unique focus-imaging mode,Photoacoustic Microscopy(PAM)can achieve high-resolution imaging of a single capillary.It is suitable for research on microcirculation system deseases.However,as a scanning-imaging technology,the time-resolution of conventional PAM is not high enough.Therefore,the imaging speed is so low that it is difficult to satisfy the requirements of real-time imaging in biological research.In addition,the larger of data acquisition,the higher the system hardware requirements,which will seriously hindered the technology evolution to the clinical application.In order to solve the problems such as slow imaging speed,large amount of data and instability caused by point-to-point mechanical scanning in conventional PAM system,a non-uniform sparse sampling scheme has been proposed to improve the time-resolution of the system and realize the fast data acquisition.By analyzing the inhomogeneity of tumor and blood vessel PAM images,a more suitable non-uniform sampling scheme is proposed,which limits the scanning points to the region of interest or principal component to further reduce the total sampling points while ensuring a certain PAM image signal-to-noise ratio.Specifically,the necessity for non-uniform sampling is discussed on tumor and blood vessel PAM images and the images are classified into two types: PAM images with concentrated region of interest and non-concentrated regions of interest.A non-uniform sampling scheme based on the edge expander theory has been proposed for PAM images with concentrated region of interest.More sampling points are focused on the region of interest.Meanwhile,a non-uniform sampling scheme based on low-rank matrix approximation has been proposed for PAM images with non-concentrated region of interest.More sampling points are focused on principal component.Both two non-uniform sampling schemes can further improve the completeness of the PAM data and reduce the number of sampling points,which will accelerate the sampling process and better recover the PAM images.In order to achieve the fast and accurate recovery of PAM images,the image recovery model has been studied.Firstly,a fast image recovery algorithm based on Go Dec is designed according to low-rank matrix approximation theory.Further,the relationship between the sparse sampling mask and image recovery algorithm is analyzed based on real PAM data,and the empirical formulas are fitted to guide the parameter selection of the image recovery algorithm.Secondly,the low-rank matrix recovery has been convex relaxation as nuclear norm minimization problem,and the optimization problem is solved in the ADMM framework.The disadvantages of the image recovery algorithm with low-rank constraint,such as poor edge preservation and artifacts,are analyzed.Finally,in order to improve the precision of the low-rank recovery algorithm and recover the details of the image more accurately,a sparse constraint which is considered as a priori information of image detail and structure is combined to form a low-rank and sparse matrix recovery problem.The multi-constrained matrix recovery problem is solved with a generalized inexact ADMM method to achieve high-resolution PAM image recovery.Based on the study of the non-uniform sparse sampling scheme and the corresponding image recovery algorithms,a fast scanning photoacoustic microscopy system has been set up.Specifically,based on the analysis of conventional optical-resolution PAM,a two-dimensional optical scanner is introduced to realize the optical scanning mode and overcome the instability caused by the mechanical scanning scheme.Besides,the sparse sampling of photoacoustic data is achieved by the reasonable control of optical scanner and the synchronous control of each part of the system without increasing the system cost.With a fast data acquisition,the time resolution is improved.Finally,with the fast scanning PAM system,real PAM experiments are designed for tumor and blood vessel targets and verify the effectiveness of the sparse sampling schemes and the image recovery algorithms.The newly built fast scanning PAM system is employed to achieve the visualization of microcirculatory vascular diseases.In order to exhaustive quantify the blood vessel property of the PAM image and provide quantitative parameters which indicate blood vessel disease,a quantification and segmantation method based on multi-scale Hessian filter is proposed.In the method,a multi-scale Hessian filter based segmentation method is proposed to obtain accurate segmentation results for multi-scale features of blood vessles.On the basis of segmented images,four quantification parameters,including vessel density,vessel length fraction,vessel diameter and fractal dimension,are calculated to quantify the location and morphology information.Further,a small area quantification method has been proposed to provide quantification color maps and realize local quantification of blood vessel.Finally,the quantification method is combined with the fast scanning PAM system to achieve blood vessel imaging and quantification to further verify the effectiveness of the system.更多还原