【作者】 李晖；

【导师】 彭玉华；

【作者基本信息】 山东大学 ， 信号与信息处理， 2009， 博士

【摘要】 图像配准技术是医学图像处理领域的一个重要的和基本的研究课题。医学图像配准技术可以将来源于不同成像设备的图像,或者不同时间利用同种成像设备得到的图像进行配准,得到更丰富的信息用于医疗诊断中。全自动医学图像配准不仅可以用于医疗诊断,还可以用于指导神经手术、放射治疗计划的制定、病灶的定位、病理变化的跟踪和治疗效果的评价等各个方面,为医生提供功能和形态的综合信息。随着图像数据量的增大,对于大多数断层扫描医学图像来说,二维、特别是三维体积数据集包含巨大的数据量,极大地增加了计算的负担,无法实时实现和临床应用,这也成为限制了现阶段配准性能较好的互信息相似性测度在配准方法中的应用。不论是刚性还是非刚性配准算法,在配准过程中,常使用多分辨率图像金字塔来进行由粗到精的搜索变换系数,提高计算效率、避免局部极小值,实现自动的更精确的配准结果。但是常见的图像金字塔,例如小波金字塔,滤波器的张量积形式使得小波变换缺乏平移和旋转不变性,这些不变性正是在图像配准中最需要的,只有具有这些不变性,才能保证从粗尺度上得到的平移、旋转和放缩参数的准确性,从而得到准确的结果。另外,由于成像系统内在和外在条件的限制,使获取的图像常常不能满足实际的需求。从软件方面来提高图像分辨率有着极大的现实意义和应用价值,而且无需增加额外的硬件设备,节省大量的费用。超分辨率技术在视频、遥感、医学和安全监控等领域都具有十分重要的作用,例如,随着高清电视(HDTV)的发展,利用超分辨率技术将大量的DTV信号转化为与HDTV接收机相匹配的信号,提高电视节目的兼容性。目前图像的超分辨率重建比较常用的是插值方法。常见的插值方法有近邻插值、双线性插值、双三次插值等,计算量比较小,但是图像插值后常常出现方块效应或细节退化(边缘模糊)。较好的视觉质量和较低的算法复杂度是处理视频、网络信号的关键要求。针对多分辨率医学图像配准和自适应图像插值技术,本文主要研究了以下几个方面:1.具有平移和旋转不变性的图像多分辨率分解可以提高配准过程的精度,避免陷入局部极值,导致误配准。提出一种改进的圆周对称多分辨率分解算法,用环形带通滤波器代替高通滤波器,使各尺度的子带都具有平移和旋转不变性。不丢失信息的前提下,降低计算的冗余度。2.改进的圆周对称塔式分解具有平移和旋转不变性,可以提高图像配准的性能。塔式分解的低频子带能量集中,用互信息进行配准;高频子带提供重要的解剖结构信息,构造基于边缘信息的模糊梯度场,并用模糊贴近度作为相似性测度,与互信息相结合。实验结果验证了新的金字塔算法的有效性,可以实现多模态医学图像的配准,结合模糊梯度场的配准算法提高了算法的鲁棒性,快速准确稳定地实现医学图像配准。3.为了提高医学图像配准的速度和有效性,提出一种结合改进的圆周对称塔式分解和边缘特性的图像配准方法。利用新的塔式分解结构中的带通子带进行处理,计算包含图像边缘信息的带通子图像的主轴和质心,作为低频子带配准过程的初值,提高配准的速度,也提高了对于发生较大形变的图像或图像初值比较大时进行配准的鲁棒性。4.针对数据量巨大的三维医学图像,提出了一种球形对称多分辨率金字塔分解框架(SSMP),McClellan算法用于实现球形和球环形的三维滤波器组。SSMP可以对三维图像实现低冗余度且具有平移和旋转不变性的多分辨率表示。对CT和PET医学图像进行的仿真实验实现了基于球对称塔式分解的医学图像配准的框架的配准过程,提高全局优化的性能,达到最优的变换参数。5.提出一种基于图像分类和非线性插值核的快速自适应插值算法。首先通过图像像素分类方法确定各像素的方向,并将像素点分为边缘区域、纹理区域和平滑区域;基于方向利用适当的提升框架中的Neville插值滤波器对各区域中的像素点进行自适应插值计算;为了得到更好的视觉效果,通过拉普拉斯算子对灰度变化大的区域进行增强处理。实验结果验证,本文算法在提高信噪比的同时,不仅减少了运算时间,在保持边缘方面也得到了较好的效果。更多还原

【Abstract】 Image registration plays a crucial role in medical image analysis by providing comparative and/or complementary information from multimodal images or images taken at different times for the same or different subjects. Automated image registration has thus been widely used in clinical applications, such as diagnosis, staging, assessment of the response to the treatment, and image guided surgery.With the advance of medical imaging techniques, the image data size increases dramatically. As a result, the computational complexity for image registration, in particular three-dimensional (3D) image registration, increases exponentially, which leads to a higher possibility of mismatching, i.e., the current intensity-based similarity measures may likely be trapped into local extrema. While multiresolution analysis (MRA), such as wavelet transforms, provides a potential mechanism to improve the registration accuracy and reduce the computational complexity, it does not have the necessary properties, such as translation- and rotation- invariance, required for image registration.In many image processing applications, it is necessary to interpolate digital images so that high resolution images can be obtained from low resolution images. For example, digital TV signals are transformed into signals that match those of high-definition television (HDTV) receivers. Image interpolation algorithms are also in demand in the field of remote sensing and sensor networking, where low resolution images are usually captured by inexpensive imaging devices.In practice, simple interpolation methods, such as nearest-neighbor, linear, bilinear, and bicubic interpolation are most widely used. However, these methods usually yield an interpolated image with blurred edges, which degrade the perceptual quality of the image. Improving the subjective quality and reducing the computational complexity of interpolation algorithms are important issues in video and network signal processing.In multiresolution medical image registration and image interpolation, the main contributions in this thesis including:1. The translation- and rotation- invariance of the multi-resolution analysis improvesthe registration accuracy and avoids trapping in local extrema which frequentlyleads to misregistration. A new multiresolution analysis, improved circular symmetric multiresolution decomposition is proposed. An annular band-pass filter takes place of the high pass filter in the circular symmetric multiresolution analysis to reduce the redundancy. All the subbands possess translation- and rotation-invariance.2. The improved circular symmetric image pyramid decomposition with translation-and rotation- invariant properties can improve the performance of image registration. Low-pass subband has noise-removing property and is suitable to image registration based on mutual information. Band-pass subband has more significant structural information, which establishes image fuzzy gradient field and constructs the fuzzy approach degree. A coarse-to-fine procedure is adopted to utilize these features to achieve registration procedure. Experiments demonstrate the good performance of the proposed novel pyramid decomposition. The local extrema can be reduced and these characteristics of combined measures yield more robust and accurate registration results.3. To improve the medical image registration accuracy and efficiency, a new approach of image registration based on circular symmetric multiresolution decomposition and edge information is proposed. Firstly low pass subbands in the multiresolution decomposition preserve the global image information, and are utilized to perform image registration based on mutual information. Then the cross-weighted moments are calculated in the first level of band pass subband which includes sufficient spatial information. It provides the initial transformation parameters to the hierarchical registration. So the transformation displacements will be calculated rapidly and accurately in the optimization algorithm. Experiments demonstrate that the intensity and edge information combined method based on circular symmetric pyramid improves the registration accuracy and robustness. It also reduces the iterations of optimization and has low computation complexity.4. A novel 3D spherical symmetric multiresolution pyramid (SSMP) is proposed. In our SSMP, McClellan algorithm is applied to build the sphere and spherical ring 3D filter banks. SSMP can be used to generate a multiresolution representation of 3D image with a low redundancy framework which possesses the translation- and rotation- invariant subbands. Our experiments on clinical CT and PET datasets have demonstrated that the registration performance, both in terms of speed and accuracy, based on this new multiresolution decomposition has been improved significantly.5. We propose a fast adaptive image interpolation algorithm that classifies pixels and uses different linear interpolation kernels that are adaptive to the class of a pixel. Pixels are classified into regions relevant to the perception of an image, either in a texture region, an edge region, or a smooth region. Image interpolation is performed with Neville filters, which can be efficiently implemented by a lifting scheme. Since linear interpolation tends to over-smooth pixels in edge regions and texture regions, we apply the Laplacian operator to enhance the pixels in those regions. The results of simulations show that the proposed algorithm not only reduces the computational complexity of the process, but also improves the visual quality of the interpolated images.更多还原