【作者】 李林；

【导师】 左焕琛； 宋志坚； 周国民； 李文生；

【作者基本信息】 复旦大学 ， 人体解剖与组织胚胎学， 2005， 博士

【摘要】 背景：脑是人体结构最为复杂的器官，准确、清晰地了解与掌握其内部各种神经结构的形态信息是对相关疾病进行诊断和治疗的重要基础。必须对脑内结构的空间形态和位置进行详细而科学的研究。过去在临床上人们一般通过CT、MRI等影像设备获取脑部的二维断面，医生根据二维图像来想象各个神经结构的空间关系，随着计算机技术的发展，计算机辅助三维重建被广泛应用于医学领域，人们开始用CT、MRI等断面图像来进行三维重建。但由于受成像原理的限制，精确度较差，所能提供的人体内部信息较少，因此其应用范围受到局限。近年来人们用脑图谱的方法来研究脑的形态结构。但从国内外研究现状来看，脑图谱的研究还存在着缺陷：有的用手工勾画而成的，带有一定的主观性；另外多数是以MRI为数据源，MRI图像分辨率较低，缺少许多人体颅内的重要形态和功能信息。虚拟人技术的出现为脑图谱的研究提供了新的契机。虚拟人数据集是以组织切片技术获得的，具有精度高、能准确显示神经解剖结构信息的优点，但其数据只代表个体的信息，如何把其中的丰富的神经解剖学信息应用到临床，为临床患者服务是目前需要迫切解决的问题。医学图像融合技术为解决这一问题提供了新的手段，目前医学图像融合技术在诊断与治疗中应用较多，但多是针对同一病人的CT、MRI、PET或SPECT等图像。尽管影像技术与设备不断提高，但对于诸多重要且临床上迫切需要的信息，这些设备均无法获得或清晰显示的，关于如何获取这些信息，目前文献还较少报道。目的：尝试解决临床诊断与治疗时影像数据中脑神经结构信息量不足的问题。方法：以中国虚拟人数据集为对象，用基于阈值的自动分割和基于解剖学知识的手工分割相结合的方法对其中部分神经结构进行识别与分割。为消除因手工分割产生的误差，在分割过程中用图像透明的方法，并用腐蚀和膨胀对分割后的神经结构进行平滑处理。用Chamfer Matching方法来配准融合MRI和PET图像，并对Chamfer Matching方法进行了改进，采用了可变步长逐步逼近的算法，使精度大大提高。用基于最大互信息方法把虚拟人数据集与病人MRI数据场进行配准融合，采用刚体配准方法，使两幅图像达到刚体位置上的配准，针对两者存在的个体差异，采用非刚体配准方法予以校正，从而达到了精确配准的目的。最后用表面绘制与体绘制相结合的方法来显示融合后的神经结构，并提出了基于等值面原理来提高体绘制的运算速度。结果：经过三维重建后的虚拟人神经结构轮廓清晰，形态逼真。将外部结构透明后可清晰显示各个内部神经结构的形态、毗邻关系及在脑中的位置。可在三维空间中绕任意轴旋转任意角度，从不同的方向进行观察。经过刚体与非刚体配准后，虚拟人数据与MRI数据基本达到对位重合，在同一幅图像上既能显示虚拟人数据集中的神经结构，又能显示病人MRI数据中的病变部位。重建后的图像逼真，可从三维上清晰显示病变部位及神经结构的形态、毗邻关系及在脑中的位置。结论：三维重建后的神经结构，对于神经外科疾病的诊断与治疗和解剖学教学科研具有重要的辅助作用。将虚拟人数据集融合到活体病人影像数据场中，能够成功解决临床影像数据中神经结构信息量不足的问题，使虚拟人数据集能够为临床应用服务。更多还原

【Abstract】 Background:Brain is the most delicate organ in human body, understanding and mastering of morphology and function of brain is crucial to the disease treatment. So the position and morphology of the neural structures in brain must be investigated completely and scientifically. At past, imaging devices such as CT or MRI is used to acquire the images of neural structures two-dimensionally. The doctors have to imagine the shape and position of the structures according to these images. As the development of computer technology, the computer aided three dimensional reconstruction is widely used in medical area. The sectional images such as CT, MRI are used to reconstruct the neural structures three dimensionally. Due to the limitation of these devices imaging principle, much of the important information in skull, especially some important deep nucleuses, cannot be acquired by these equipments. Thus, their applications in clinic practice are restricted. To solve this problem, some people try the atlas, i.e., superimpose the atlas on the images after registration of the two images. But it has shortcoming: most atlases are drawn by hand and is subjected to individuals; some atlases are based on MRI images. So, the amount of information is limited by the resource data. The visible human dataset is acquired by postmortem cryosectioning technology, has the advantage of high accuracy and large amount of information, can display neural structures accurately. But it only express common information of human body, and has little value in clinical practice. Medical fusion technology provides new method to solve the problem. Recently, there are many reports about medical image fusion. But all those fusions are aimed at CT, MRI, PET and SPECT of the same patient. Although the technology and devices of imaging is improved gradually, as to much important clinical information,cannot be acquired or displayed clearly by these devices. It is seldom reported as to how to acquire these important information at present.Objective:Try to solve the problem of information shortage of radiological data in clinical diagnosis and treatment.Methods:In this research, Chinese Virtual Human dataset was segmented using automatic segmentation based on gray threshold and manual segmentation based on anatomic knowledge. In order to remove the error result from the manual segmentation, image-transparency method was used at first, then each image was smoothed by eroding and dilating, nuclear structures are reconstructed and displayed using surface rendering and volume rendering lastly. Circumscribing-circle theory is raised to improve the reconstrion speed.We registered and fused the CVH dataset with MRI. In the process of registration, Chamfer Matching method and Maximization of Mutual Information method were used separately. A changeable-step method was raised to improve the Chamfer Matching. In the registration of Mutual information, rigid registration was first used to make the two images coincide basically. Aimed at the difference between individuals, the nonrigid registration was used to revise it. Thus, the aim of accurately registration was achieved, and laid a foundation for the problem of neural structures shortage in patients’ radiological data.Results:The reconstructed nuclear structures are smooth, natural and realistic. Their shapes and positions are clearly displayed after the surface of the brain is set as transparency, and can be rotated, observed in any directions. After rigid and nonrigid registration the CVH and MRI images coincide with each other basically. The neural structures in visible human data and pathological area in MRI can be shown at the same image. After reconstruction, pathological area and neural structures can be clearly displayed three dimensionally.Conclusion:The reconstructed structures have great reference value to the clinical diagnosis and treatment and anatomic teaching and learning. Our research indicates that the fusion of the CVH and patient’s MRI can solve the problem of shortage of neural information and lay a foundation for the clinical use of CVH dataset.更多还原