【作者】 汤敏；

【导师】 王惠南；

【作者基本信息】 南京航空航天大学 ， 精密仪器及机械， 2007， 博士

【摘要】 医学图像处理及应用,是计算机组合诊疗系统中的关键技术,也是一个多学科交叉融合的研究领域,涉及数字图像处理、计算机图形学以及医学领域的相关知识。在目前临床应用中,大多数医生还是借助CT、MRI拍摄的二维断层图像,根据临床经验进行局部分析和诊断,这对于病变区域的精确定位和定量分析是远远不够的。因此,在普通计算机上实现医学图像处理,将使临床医生可以对可视化三维图像任意剖分。这对于诊断医学、手术规划、放射治疗规划、模拟仿真及解剖教学等方面,都具有重要的学术意义和应用价值。本研究课题是南京航空航天大学“十五”重点学科建设项目“计算机集成诊断与治疗系统”中的子课题。论文将详细阐述作者攻读博士学位期间在医学图像处理方面从事的研究工作,主要涉及图像分割、图像三维重建及可视化、图像测量这三个关键技术的研究,包括各种基础算法的理论研究和应用系统的设计研发。论文主要研究工作及创新成果归纳如下:1)针对图像分割算法的具体应用,在分析综合融合变形模型、小波变换、数学形态学等先进理论的基础上,将新型图像分割算法——梯度向量流场的变形模型和改进的分水岭算法应用于颅脑内胼胝体和脑部星形细胞瘤的分割提取,取得了良好效果。针对彩色视网膜血管图像,提出一种基于网格划分的全自动分割提取算法,保证眼底测量结果的准确性、客观性和实用性。2)针对数字化医疗系统的需要,利用普通配置计算机群和医院内部局域网进行超大规模医学数据并行重建的研究,是符合我国大多数医院实情的有效手段,也是医院PACS系统建设的必然要求。该方法较好地解决了运算速度、内存空间以及资金投入这三大难题,既节省大量资金,又能够满足临床诊断需求。3)将分形和分形维数的基本方法引入医学图像定量测量中来,首次将分形维数用于定量考察和分析不同浓度神经生长素对体外培养新生大鼠背根神经突起生长的影响。与采用显微镜进行定性观察相比,该方法不仅可以测量复杂结构的变化,而且可以描述它的生长和演变过程。分形维数的测量方法能够从量化角度更准确深刻地说明神经生长素能够促进大鼠背根神经节神经突起的生长,而且生长状况表现出与剂量的依赖关系。4)针对生物医学微细结构研究的需要,进行激光扫描共聚焦显微镜图像计算机处理的相关研究。首先对显微镜图像进行激光吸收和散射的补偿以及基于最大后验的图像盲解卷积,改善图像模糊,提高图像质量。然后利用光线跟踪算法重建三维物体,实现三维图像的平移、旋转、缩放、光线设置等操作,从而清晰全面地显示三维微细结构。此外,还可以通过电影放映方式显示三维结构随时间变化的过程,实现四维成像,为生物医学显微图像的分析提供技术支持。5)将生物医学可视化技术推进到微观切片水平,实现生物组织连续切片图像的两步配准和快速重建。考虑到生物体在序列切片图像上的重心位置具有连续性的事实,提出一种新的逐步求精的自动配准方法。在使用力矩主轴法对两幅图像进行粗略配准的基础上,对已有配准结果作微小干扰,当图像互信息最大时获得最终精确配准结果。采用基于有序体数据结构的Shear-Warp算法重建生物体内部结构,提高对体数据的遍历效率,减少对无效数据的访问,加速绘制过程。6)在充分吸收国外知名图像处理软件各项优点的基础上,利用IDL这一开发平台,进行医学图像处理分析系统MIPAS和CTA/MRA图像后处理软件VICAAT的研发,形成具有自主知识产权的软件系统。通过对实际医学图像的处理,证明该系统的理论意义和应用价值,取得大量实验结果,为医学图像处理和分析创造良好条件,并将逐步推广应用于临床诊断。归纳而言,本论文针对计算机组合诊疗系统中的关键技术,在理论上提出了一些新算法,在应用中创建了一些实用的新方法,并获得了很好的验证。同时,还研发了相应的基础应用软件,为生物医学图像处理提供了重要工具。更多还原

【Abstract】 Medical image processing is a multi-disciplinary subject, which relates to the subjects of digital image processing, computer graphics and some related knowledge of medicine. However, until now most of the clinical physicians can only utilize their experience to analyze the two-dimensional image series obtained by CT and MRI, which might be very limited to determine the perfect spatial location and exact area of the disease. Therefore, there is a crying need to build suitable software system based on common-configuration personal computer, which can be applied to visualize the corresponding three-dimensional images by image reconstruction techniques. Furthermore, through the necessary splitting and some other useful analyses or simulations before the actual operation, the clinical physicians could be more confident about the disease. In brief, study on processing and analysis of medical images is widely used in diagnostic, surgery and radiotherapy planning, and teaching in anatomy, which is of great important significance on science and worthiness in practical applications.The research project is one part of“The Diagnostic and Therapeutic System of Computer Integration”, supported by“The Tenth Five Years Constructive Finance”in College of Automation Engineering, NUAA. This dissertation describes the author’s work on medical image processing and analysis, including algorithm research and software development. The research focuses on three important research fields: image segmentation, reconstruction and measurement. The main research fields and innovative results in this dissertation focus on the following aspects:1) Two novel segmentation algorithms: gradient vector flow deformable model and improved watershed algorithm, are implemented and applied to segment corpus callosum and astrocytoma, both achieved favorable results. Another automatic segmentation algorithm for color retinal vascular images, whose performance exceeds traditional methods, is proposed to guarantee accuracy, objectivity and practicability.2) On the basis of grid computing, the essence and key techniques of parallel visualization of large medical datasets are discussed based on Intranet and common-configuration computers of hospitals. It is suitable and effective for our countries’most hospitals and is also the outcome of PACS construction. It is demonstrated that this method provide promising and real-time results, which resolve the computational speed, memory requirements and undercapitalization puzzles. 3) The neurite growth of cultured dorsal root ganglion is detected by fluorescent immunocytochemistry treated with nerve regeneration factor in different concentration. A novel method based on triangular prism surface area is introduced to calculate the fractal dimension of the two-dimensional immunofluorescent images. Experimental results demonstrate that this method is easy to understand and convenient to operate, with quantitative results according to observations of microscope.4) The methodology is proposed for image processing and interactive visualization of laser scanning confocal microscopy datasets, including automatic pre-processing and ray-casting reconstruction. The program allows for convenient, fast, interactive examination of unknown cell structures even on common PC, which significantly improve the task of understanding the internal structure of laser scanning confocal microscopy image stacks. Meanwhile, it can be used to visualize dynamically changing temporary structures conveniently in addition to static images.5) A novel method, two-step slices registration and fast Shear-Warp reconstruction, is proposed for serial tissue section images. Based on the elementary result of the principal axes transformation method, the optimal registration result is achieved when the mutual information reaches maximum. An improved Shear-Warp algorithm based on sorted volumetric data structure is applied to reduce the access time of non-contribution data cells and consequently speed up the reconstruction process.6) Several famous foreign image processing software are analyzed and evaluated respectively, and advantages are derived and absorbed from them. Two software systems: medical image processing and analysis system (MIPAS) and vascular image computer assisted analysis tool (VICAAT) are developed based on IDL language, with some examples demonstrated. Modular programming is used to design these systems, which are easier to be modeled, organized, maintained and extended by using encapsulation, inheritance and polymorphism features of objects.The new algorithms and methods built in this dissertation are of important value to theoretical research and clinical application, along with the corresponding software, which provide important tools for medical image processing and analysis.更多还原