【作者】 金晶；

【导师】 沈毅；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2008， 博士

【摘要】 作为医学图像处理与分析中的基础及关键技术,医学图像配准具有重要的临床应用价值,不仅可以用于病症性状的诊断,还可以通过对病灶部位进行跟踪来引导治疗过程以及对治疗效果做出评价。在图像融合、三维图像重建以及外科手术导航中,广泛应用目前医学图像配准方法来定位图像之间的空间位置关系。本文以医学图像配准研究为背景,对刚体图像、序列刚体图像和非刚体图像的配准方法进行了研究,并针对序列图像以及非刚体图像配准需要处理的数据量大的问题,着重研究了全局化的参数寻优方法以及多分辨率分析的配准策略。本文的主要内容包括:应用非线性相关度量结合下山式单纯形搜索方法对医学图像进行了配准研究,提出了最大化非线性相关系数的刚体图像配准方法。由于非线性相关系数是对互信息的改进,它能够以[0,1]之间的数据量化两个变量之间信息相关的程度,因此比互信息更易于对图像之间的相关程度进行比较和分析,描述更加直观。利用非线性相关系数的极值性对下山式单纯形方法进行改进,应用代价函数的上下门限来限定优化算法对于极值的搜索过程,从而克服非线性相关系数的非线性所引起的局部极值问题。变容许门限的引入可以减少局部极值处的搜索迭代次数进而提高算法的快速性。仿真验证了该方法的性能不受浮动图像和参考图像之间对比度差异的影响,而且可以应用到多种成像模式的配准中。针对序列图像不同内部关系的特点,提出了两种新的序列刚体图像配准测度。对于具有已知内部关系的序列刚体图像配准,其配准模型是首先选择序列图像中的第一幅作为浮动图像与参考图像进行配准,然后取第二幅图像作为浮动图像与前两幅已配准的图像同时进行配准,依此类推。针对于该配准模型,提出一种新的相似性测度——链式互信息作为衡量序列图像是否配准的标准。该测度可以在前n? 1幅图像精确配准的基础上,敏感第n幅待配准图像与前n? 1幅图像之间的相关性大小。实验证明,该测度可以很好地配准具有已知内部关系的序列刚体图像,其配准结果具有亚像素的精度。同时,与两幅图像的配准结果相比较,进一步验证了利用最大化链式互信息进行已知内部关系的序列图像配准的性能优势。对于具有未知内部关系的序列刚体图像配准,提出使用非线性相关信息熵作为配准测度。非线性相关信息熵以[0,1]之间的数值具体量化多个变量之间的普遍关系,而且不受序列图像位置顺序的影响。仿真从旋转和平移两个角度对非线性相关信息熵的序列刚体图像配准进行了性能评估,其结果验证了非线性相关信息熵敏感具有未知内部关系的序列图像相关关系的有效性。在非刚性图像配准的框架下,提出应用Wendland紧支径向基函数作为参数化的空间变换模型,并针对非刚性配准模型的未知参数量大的问题,提出应用全局化的粒子群优化算法对整个可行性空间进行快速而准确的探索和挖掘,以求得非刚性变形的未知参数。粒子群算法具有快速全局寻优的特性,但是由于粒子容易陷入早熟收敛状态而使其应用受到限制。本文对基本的粒子群算法进行了改进,提出了基于变邻域选择的粒子群算法。该方法在陷入局部最优粒子的邻域内重新选择优于该局部最优粒子的参考粒子来更新早熟粒子,使其能够从早熟收敛的状态跳出,进而继续进行全局最优值搜索。实验证明,改进的粒子群算法结合Wendland紧支径向基函数的非刚性图像配准方法,可以有效地对具有全局以及局部非刚性形变的图像进行配准。针对图像配准的速度和质量要求,提出了基于整数提升小波变换的多分辨率配准模型。该配准模型首先对待配准图像进行多分辨率分解,然后取其近似图像进行配准得到变形参数,再将配准后的分解图像重构。由于整数提升小波变换可以实现整数到整数的变换,可以对图像实现无损重构,因此比第一代小波变换更利于对图像进行多分辨率分析。实验中利用正交小波以及双正交小波分别对超声图像进行了多分辨率分析,其结果证明,整数提升小波变换的多分辨率分析,可以使分析后的结果图像保留更多的原始图像信息。同时应用双正交小波的整数提升变换对超声图像进行不同分解层次下的配准,验证了整数提升小波在多分辨率分析配准中可以有效地减少算法的运算迭代次数及计算时间。利用本文提出的刚性图像配准方法、非刚性图像配准方法,并结合多分辨率分析对肾部超声造影图像进行了配准。通过对配准后的图像序列进行时间—强度曲线分析,可以得出病灶部位良、恶性的正确结论。该应用同时证明了本文对于医学图像配准算法的研究具有一定的临床诊断意义及实际应用价值。更多还原

【Abstract】 As a key technology of medical image processing and analysis, image registration is very important in clinical applications. It not only can be applied in the diagnosis of disease, but also can help in clinical operations by tracking the focus part and in accessing the treatment. Nowadays, medical image registration is widely applied in image fusion, three-dimensional image reconstruction and surgical navigation to locate images.This dissertation mainly focuses on the registration of medical images, especially the rigid images, series rigid images and non-rigid images. As far as the processing of very large volume data in series and nonrigid image registration applications is concerned, the global optimization method and multi-resolution analysis based registration scheme are also discussesed. The main contents of the dissertation are as follows:Applying the nonlinear correlation metrics and downhill simplex searching technique, the dissertation proposes the rigid image registration method based on the maximization of Nonlinear Correlation Coefficient (NCC). As an improved version of Mutual Information (MI), NCC can quantitatively describe the nonlinear correlation degree between two variables using a value in the closed interval [0, 1]. Therefore, it is more intuitionistic and suitable to compare and analyze the correlation degree among images. The extremum of NCC can help us to adopt the upper and lower thresholds of the cost function to revise the downhill searching process for the optimal, and overcome the local minimal problem induced by the nonlinearity of NCC. The introduction of variant accuracy tolerance will reduce the iterations at the local minima in the searching process, and then, enhance the speed of the algorithm. Simulations verify that the proposed method can be applied in multi-modal image registration and its performance will not be affected by the contrast differences between the floating and reference images.According to the different characteristics of the inner geometry connection of series image, the dissertation proposes two types of metrics to register the series rigid image. For series image with known inner geometry connection, its registration can be carried out by selecting the first image in the series as the floating image, which will be registered to the reference image. The second image in the series is registered to the reference image and the registered first image simultaneously. The rest of the series image may be deduced by analogy. A new correlation metric, Shared Chain Mutual Information (SCMI), is proposed for this registration model to ensure the series image to be accurately registered. SCMI can quantitatively describe the correlation degree between the nth image with the pre-registered n-1 images. Experiments have shown that SCMI can be used to register series rigid image with known inner geometry connections, and can achieve sub-pixel registration accuracy. Comparisons with the accuracy of two image registrations further prove the advantages of SCMI on registering the series image. For registration of unknown-inner geometry connections, Nonlinear Correlation Information Entropy (NCIE) is proposed as the registration metric, which can use a value in the closed interval [0, 1] to estimate the general relationship among multi-variables, and may not be affected by the order of images. Simulations on the images with rotation and translate transformations have been conducted to verify the performance of NCIE as a registration metric, and the results prove the effectiveness of NCIE to series rigid image with unknown-inner geometry connections.Under the nonrigid image registration frame, the dissertation proposes to use Wendland compactly support radial basis function as a parameterized transformation model. As the parameters are numerous in non-rigid registration model, the particle swarm optimization algorithm is selected to achieve exploration and exploitation in the whole feasible space accurately and fleetly for the unknown parameters of nonrigid transformation. Particle swarm optimization algorithm has the characteristic of fast global searching for the optimal, but it is restricted in applications by the inclination of its particles falling into premature convergence. The dissertation improves the basic particle swarm optimization algorithm and proposes the revised Variable-Neighborhood-Selection based Particle Swarm Optimization (VNS-PSO) algorithm. The algorithm updates the premature particles by re-assigning a better reference particle in its neighborhood. This will lead the algorithm out of the premature state and continue its global optimization searching. Experiments show that, the registration method combining the revised VNS-PSO algorithm and Wendland compactly support radial basis function can effectively register the images with global or local nonrigid transformations.To meet the speed and quality requirements for series and nonrigid image registration, the dissertation proposes an Integer Lifting Wavelet Transform (ILWT) based multi-resolution analysis registration model. This model decomposes the images firstly, registers the approximate images to obtain the transformation model secondly, and finally, reconstructs the registered image by applying the transformation parameters to the original resolution image. Comparing to the first generation wavelet transformation, ILWT can implement transformation from integer to integer, and lossless reconstruction of image. Therefore it is more suitable for multi-resolution analysis than the first generation wavelet transformation. In the experiment, orthogonal wavelet and biorthogonal wavelet are used to decompose ultrasonic image multi-dimensionally. Results show that ILWT based multi-resolution analysis can keep more original information in the processed images. Moreover, the registration results of the ultrasonic images decomposed at different resolution level by biorthogonal wavelet verify that ILWT can effectively reduce the iteration and the calculation time of the algorithm.Using the proposed rigid and non-rigid image registration methods and multi-resolution analysis strategy, the kidney ultrasonic series image is registered. By analyzing the time-intensity curves of the registered series image, we can achieve the correct conclusion about whether the focus is benign or malign. This application further validates that the researches on medical image registration methods in this dissertation are applicable in clinical diagnosis.更多还原