节点文献

医学高维数据的临床环境实时计算研究

Real-time Computing of Dynamic Volume Data in the Clinical Environment

【作者】 郝立巍

【导师】 陈武凡;

【作者基本信息】 第一军医大学 , 生物医学工程, 2007, 博士

【摘要】 近年来,现代大型医疗成像技术有了飞速地发展,医学影像在临床诊断与医学研究中扮演了愈来愈重要的地位。医学影像呈现出实时化、海量化的趋势。海量信息经过有效地整合、加工与挖掘,使医学影像数字化分析技术逐渐从临床诊断的辅助手段,成为实现精确诊断的必需。所以,随着临床诊断要求的提高以及医学影像学研究的开展,迫切需要更高效的医学图像数字化分析技术,对医学图像数字化分析技术的研究刻不容缓。这主要表现为以下两个方面。首先,新的成像技术催生新的数字化分析技术。例如,现在的动态三维(4D)超声能让医生和母亲看到未出生胎儿的脸部特征和在子宫内的运动。这是因为三维超声能探测、采样、数字化存储三维的超声回声信号,并能渲染绘制出活灵活现的未出生胎儿的图像。四维超声还能实时地动态接收和显示这种三维图像,帮助医生研究胎儿的移动与活动规律。对临床医生来说,利用四维超声能观察胎儿健康状况和微小的移动,可以在屏幕上对未出生的胎儿做从头到脚的健康评估,就象儿科医生直接面对已出生的婴儿一样。通过这样对未出生婴儿的体位变化以及呼吸的观察,医生能更容易诊断。其次,高分辨率的图像、很短的成像时间以及日益广泛的临床应用使后处理研究成为必需。例如,多排CT的成像不再受患者屏住呼吸的时间间隔的限制,使人体心脏成像和周边血管成像成为可能。与传统单排CT相比,多排CT一次成像的层厚更薄、层数更多,所以细节分辨率更好而且数据量大,因此放射科医生在高级计算机后处理软件的辅助下才能更好更快地从海量的信息中提取更多有用的诊断信息,而不致于耽误或遗漏。另外一个例子是CT仿真结肠镜,与有创的传统光学结肠镜相比,它的检查时间较短,患者耐受性较好。CT仿真结肠镜可以清晰地显示结直肠内的病灶,并可以从不同角度观察其周边解剖结构。为此,本研究主要是针对临床环境的多维数据的实时计算作了有针对性的研究,其主要内容如下:第一.基于通用微机平台图形处理器(GPU)的快速医学图像处理研究。伴随着微机的功能升级,计算机图形学的研究与应用发生了从工作站向微机的大规模转移。这很大程度上是由于微机平台图形处理硬件的发展和革新。近年来,随着微机平台图形处理器性能的大幅度提高以及可编程特性的发展,人们开始将图形流水线的某些处理阶段以及某些图形算法从CPU向GPU转移。除了计算机图形学本身的应用,还包括其他领域的计算,通用计算已成为GPU近年来的研究热点。本研究从图形硬件发展的历史开始,重点探讨了GPU在医学图像处理领域的应用、其技术原理和发展状况,并进行了比较研究。第二.基于的GPU加速的Gibbs随机场分割算法研究。图像分割是指将图像划分成一系列彼此互不交叠的匀质区域。它是图像处理和计算机视觉最基本问题之一,是实现从图像处理到图像分析,进而完成图像理解的关键性步骤。基于吉伯斯随机场的先验模型是解决退化图像病态逆问题正则化求解的重要理论模型。它通过提供良好的空间上下文约束信息,在贝叶斯医学图像分割中运用广泛。然而,在临床分割实践中,由于计算速度问题,需要适当改进以适应临床的不同需求。本文正是在这种背景下,提出了基于GPU加速的Gibbs随机场分割算法的改进:算法改进的加速性在于用GPU中的片元作色器并行执行分割算法中的的逐点迭代计算,取代CPU串行执行的逐点计算,大大减小了CPU的计算负载,效率大大高于单独采用CPU的分割算法,使基于Gibbs随机场的模糊C均值分割算法运算接近实时,大大提高了Gibbs随机场分割算法在临床的实用性。第三.提出了基于Gibbs形态学梯度的标记点分水岭分割算法的骨自动分割以及基于轮廓一致性的分割验证。利用序列CT图像对骨组织做自动分割是计算机辅助骨科的重要技术。由于骨结构的不一致性、病理改变以及CT数据内在的模糊性,完全自动的分割面临很大的困难。本文的研究目的是提供一种有效的解决框架,既避免了传统分割中大量的人力介入,又避免了自动分割造成的错误。具体步骤是:设计合适的自动分割算法对序列CT图像中骨组织做自动分割;随后利用相邻CT图像上骨轮廓是渐进变化的特性,通过比较提取的轮廓形状的一致性来自动验证分割的结果;对一致性较差的结果引入决策判别机制。目前,计算机辅助外科研究是生物医学工程的一个研究热点。本文针对该研究中所面临的二维图像区域轮廓线自动提取及轮廓验证问题进行了探索,并提出了有效的解决框架。通过实验表明,本文提出的解决框架能自动地准确分割CT序列图像中的骨组织,并对分割失败的图像能通过相邻轮廓相似性比较自动检测出,并交由根据实际需要设定的人工干预或其它判别方法特殊处理。这样既保证了分割结果的准确性,又极大减少了计算机辅助外科中的人工干预,对计算机辅助外科的推广能起到一定的作用。第四.提出了基于GPU硬件加速的动态三维面绘制研究。等值面重建被广泛应用于标量场可视化。特别是在医学领域,大量的医学研究和临床观察需要借助于医学体数据的表面重建结果。本研究提出了一种加速算法:这种加速算法利用了普通微机平台上的图形处理器(GPU)中的定点作色器(Vertex Shader)的可编程性,把等值面计算的巨大计算量从CPU转到了GPU,从而极大加快了重建速度。这项研究能在普通微机体系上就能实现动态三维数据的实时面绘制计算,因为算法无需一次性把三维动态体数据全部装载入计算机内存,也无需在内存中保存三维动态体数据的每帧画面的重建结果,更无需在内存中保存三维交互后,每帧等值面的计算结果,所以完整的三维动态体数据可以保存在硬盘上,根据需要在内存中装载对应帧的数据,在普通微机上就能完成相应的计算。利用本算法的缺点是不能保存等值面重建的结果数据。但是,在实际开发过程中可以发现,本算法更适合与体绘制的混合绘制及人机交互。第五.提出了基于GPU硬件加速的动态三维超声实时体绘制研究。三维动态超声具有非常广泛的发展前景和临床应用。但是三维动态超声的数据结构不属于三维规则数据场,所以不能直接在微机上采用三维纹理硬件加速算法实现实时体绘制。本文首先介绍了基于微机平台,针对三维规则数据场的三维纹理硬件体绘制加速算法。通过对现代微机显卡结构的分析讨论,本文提出了利用现代微机显卡中的顶点渲染器编程实现超声的三维动态扇扫数据场到三维规则数据场的快速变换,然后利用显卡中的三维纹理快速运算功能,从而在普通微机平台上实现了三维动态超声的实时体绘制。本方法具有很广泛的应用价值。在试验平台配置的计算机性能几乎相同和试验数据大小几乎一致的情况下,本算法的优势在于:1)绘制速度提高了将近10倍,已接近实时;2)绘制速度受观察角度的影响较小;3)不需要事先规整三维超声数据;4)在绘制时无需在内存中保留3份数据拷贝。伴随着成像技术的发展,医学影像数字化分析技术已从临床诊断的辅助手段,成为实现精确诊断必需的过程。为此,本论文研究的意义如下:提高诊断质量的需要:成像设备采集的海量影像信息,如不整合、加工与处理,信息仍然是孤立的、低质的,如此繁杂无序的信息往往使得临床医生感觉盲然无措,这不利于临床诊断,不利于病因的有效与定量分析,不利于对疾病的新发展与新认识。目前,医学影像后处理的研究已成为国内外研究的热点。但是目前的分析方法与临床的诊断的要求还有一定的差距,分析的精度与时间上还需提升,研究对象有待扩宽。若想进一步提高诊断质量,有赖于诸如图像分割及其可视化等技术的发展。符合医院数字化建设的需要:数字化医院是将最先进的IT技术充分应用于医疗。其核心是通过宽带网络把数字化医疗设备、数字化医学影像系统和数字化医疗信息系统等全部临床作业过程纳入到数字化网络中,实现临床作业的无纸化和无片化运行。医学图像通信与归档系统(PACS)是数字化医院的主要组成之一,PACS系统完成了医学影像的存储、通信功能,从而实现了数字影像的共享,打破了数字影像只由影像科独占的局面,但当前医学影像的数字化分析技术并没有实现共享,临床医生在PACS看到的只是由设备采集的原始二维图像数据,若欲进行进一步的分析,需要到影像科室由设备厂家提供的为数不多工作站上进行,而这些工作站的使用权又往往集中在少数医生手中,这严重影响了医学影像数字化分析技术在临床诊断中作用的发挥。本研究正是为了有效解决这一需求,从而推动医院的数字化建设。本文在这个领域做了一些有益的探索,为最终针对临床的高维医学图像分析处理系统打下了研究基础。

【Abstract】 Recent years, because of fast development of modern medical imagingequipments, they are playing more and more important roles in clinical diagnosis andmedicine research. Multi-dimensioned real-time screening of medical imaging isbecoming popular worldwide. Thus high volume medical information must beprocessed, integrated and mined effectively before clinical diagnosis. Intelligentanalysis of multi-dimensioned medical data is becoming necessary. The demand ofclinical diagnosis and research of medical radiology raise appeal for high-throuputanalysis methods of medical images, which is shown in two ways.Firstly, new imaging technology creates new digital analysis in medicine. Forexample, now four-dimensional (4D) ultrasound lets women and doctors look atfacial features and watch the growing baby move. With 3D ultrasound, a volume ofechoes is taken, stored digitally, and shaded to produce life-like images of the fetus. A4D ultrasound takes the images produced by 3D ultrasound and adds the element ofmovement. Now, the life-like pictures can move and the activity of the fetus can bestudied. To doctors, 4D reveals more detail about fetal health and small movement.Just as a pediatrician begins an exam by observing a newborn, doctors assess the fetusfrom head to toe on screen. Watching him or her shift position and breathe, doctor cancheck for problems.Secondly, high resolution images, ultra-fast scanning speed and a broad range ofclinical applications make postprocessing study necessary. For example, multisliceCT is no longer constrained by a patient’s limited breath-hold time, allowing imagingof the heart and peripheral vessels. With a combination of more slices and thinner slices, multislice CT captures significantly larger data sets and yields sharper, moredetailed images. Thus it is useful to utilize sophisticated computer technology toenable radiologists to capture enormous value from advanced multislice images. With3D, radiologists can review large image sets quickly and easily for importantdiagnostic information that might otherwise be missed. Another example is VirtualCT Colonoscopy, which provides a comprehensive assessment of the inner colon aswell as the surrounding anatomy in one fast and easy exam compared to the invasiveoptical colonoscopy.For this purpose, the thesis mainly focuses on real-time computation ofmulti-dimensioned medical data in the clinical environment. Main topics are listedbelow.1. Study of medical image processing based on Graphics Processing Unit (GPU)is presented. Along with the fast improvement of personal computer, lots ofapplications in computer graphics are migrating from workstation to normal pc,which is mainly promoted by fast innovation of graphics hardware in pc platform.Except to applications of graphics itself, GPU is used as general-purpose processingunit in many other fields, which comes up very hot in research these years. Combinedwith a history introduction of GPU evolution in pc platform, a discussion is given toprobe into the development of GPU in medical image processing.2. Image segmentation means to split an image into a lot of homogeneousregions which are separable each other. It is a fundamental problem in imageprocessing and computer vision and is a key step to image analysis and even to imageunderstand with numerous applications.The Gibbs random fields model is animportant theory in solving the ill-posed inverse problem that needs properregularization in a degraded image, which has widely applied to medical Bayesiansegmentation due to providing an excellent spatial contextual constraints information.However, the classical GRF models must be revised because of computing speed inclinical environment. Thus, in the paper, an improved C-means segment methodbased on Gibbs random field accelerated by GPU is proposed.3. Automated bone segmentation by marker-controlled watershed based onGibbs morphological gradient and validation of segmentation using contourcoherency is proposed. Automated bone segmentation of CT image sequences is a key technology in computer-aided surgery. Elements such as the inhomogeneous ofbone, pathologies, and the inherent blurring of CT images all lead to difficulties ofcompletely automated bone segmentation. In this paper, an effective solution ispresented, which not only saves time-consuming human interaction, but also avoidsfatal errors caused by automated segmentation. Firstly, a new automated segmentationis implemented to extract bone contours. Secondly, wrong contours are detected byvalidation of segmentation using bone contour coherency in neighbor CT images.Some judgment mechanism can be adopted to reanalysis wrong contours.4. Real-time marching-cubes algorithm accelerated by GPU hardware isproposed. Isocontouring is widely used in the visualization of scalar data. Especiallyin medicine, computation of isocontours has huge applications in visualization ofsurfaces from medical volume data. An acceleration approach for renderingisosurfaces of a scalar field is presented. Using the Vertex Programming capability ofcommodity graphics cards, the cost of computing an isosurface from the CentralProcessing Unit (CPU), running the main application, is transfered to the GraphicsProcessing Unit (GPU), rendering the images.By the study, surfaces of time-varyingdatasets at distinguished threshold values can be extracted in real-time in clinicalexamination environment.5. Real-time volume rendering algorithm of dynamic 3D ultrasound acceleratedby GPU hardware is proposed. Dynamic 3D ultrasound is a very promisingtechnology in clinical use. But 3D ultrasound data can not be visualized directly withreal-time volume rendering accelerated by 3D texture hardware, because it is not aCartesian 3D dataset. In this paper, fast volume rendering using 3D texture hardwareacceleration is briefly introduced at first. Then by analyzing modern pc graphics cardarchitecture, a modified method is proposed to obtain real-time volume rendering ofdynamic 3D ultrasound data by Programming Vertex Shader in GPU. Experimentalresult shows that 3D ultrasound data can be rendered with real-time speed in normalpc platform. The method can be applied wide in future 3D dynamic ultrasoundclinical research.In all, with fast development of modern medical imaging equipments, intelligentanalysis of multi-dimensioned medical data is becoming more and more importantpart in clinical diagnosises. It is wide recognized that new medical imaging equipments can only be fully utilized by more and more study of new post-processingalgorithms and visualization methods of of medical imaging data. Some usefulmethods are studied in the paper.

节点文献中: 

本文链接的文献网络图示:

本文的引文网络