【作者】 王鹏；

【导师】 须文波；

【作者基本信息】 江南大学 ， 计算机应用技术， 2008， 硕士

【摘要】 随着数字医疗、远程诊断技术的实施与快速发展,医疗图像处理越来越受到人们的广泛关注。与普通图像相比,医疗图像本质上具有模糊性和不均匀的特点,医学上采用不同的成像设备来获取多种不同模态的图像信息,临床上需要将这些信息结合起来,为医生诊断和手术引导上提供更多的信息,这就需要将几种不同模态的图像进行融合。图像配准是图像融合的前提和基础,只有经过良好配准后的图像才能进行有效的融合。群智能算法是一种模拟自然界中社会性生物的群体行为的随机优化算法,具有全局优化性、鲁棒、并行等优点,作为图像配准中的优化策略,取得了较好的配准效果。但由于群智能算法需要进行多步迭代随机搜索,导致在全局收敛和实时性方面性能较差;另外,目前的医疗图像多为三维图像,数据量大,处理起来需要很长时间。因此,很有必要设计高效的并行优化算法以提高医疗图像配准的精度和实时性。首先,本文在分析遗传算法、蚁群算法、粒子群算法及具有量子行为的粒子群算法(Quantum-behaved Particle Swarm Optimization, QPSO)的基础上,根据群智能算法的内在并行特性,设计了适合并行执行的种群结构—动态邻域拓扑结构;并结合智能与高性能计算实验室曙光SMP机群并行服务器和联想PC集群的系统结构,设计和实现了三种类型的基于动态邻域拓扑结构的并行计算模型。基于这三种并行模型,分别采用MPI、OpenMP以及MPI+OpenMP混合编程实现了并行遗传算法、并行粒子群优化算法以及具有量子行为的并行粒子群优化算法。并将这三类算法应用于求解两类非线性优化问题-无约束最优化问题和约束优化问题,结果显示并行QPSO算法在求解多极值优化问题中比串行的群智能算法具有更强的优化能力。其次,论文研究了基于最大互信息的图像配准算法和Powell算法,并首次将QPSO算法以及基于邻域拓扑结构的并行QPSO算法作为优化策略用于图像配准。通过对MR和CT图像的多模配准仿真实验证明了该并行算法在图像配准上具有很好的性能。最后,针对图像配准中的两个关键步骤进行了改进:引入了新的相似性测度;提出了基于并行QPSO算法和Powell算法的混合搜索优化策略。在此基础上,应用新的相似性测度和混合优化算法完成了2D to 2D的单模态图像的配准,结果表明基于混合搜索优化策略算法的图像配准达到了亚像素水平,且算法具有较好的实时性。更多还原

【Abstract】 With rapid increasing of Digital Medical and Remote Diagnosis technology, Medical Image Processing has caused wide attention. Compared with common images, Medical image has the characteristics of fuzzy and not uniform. Several different modals image information is acquired by various imaging device. For providing more diagnosis information to the operation instructs and the doctor, this information should be integrated into one image in clinical and the Image Fusion technology is needed. Image Registration is the precondition and foundation of Image Fusion. The image can be fused efficiently after the image has been registered accurately.Swarm Intelligence (SI) is the property of a system whereby the collective behaviors of agents interacting locally with their environment. SI algorithm is a kind of stochastic search method that can solve the specified problems by simulating the collective behaviors. The characteristic of SI algorithm is stochastic, parallel, robust and distributed. As a kind of optimization strategy in Image Registration, SI algorithms have got good effects on the image registration. However, SI algorithms have poor performance on the global convergence speed and real-time property as they need lots of iterations to complete the search. In addition, most medical images are three-dimensional and have a large amount of data to be processed. Therefore, high performance parallel optimization algorithm for medical image registration is needed.Firstly, based on the analysis of Genetic Algorithm (GA), Ant Colony Optimization (ACO) Algorithm,Particle Swarm Optimization (PSO) Algorithm and Quantum-behaved Particle Swarm Optimization (QPSO) Algorithm, a dynamic neighborhood topology model (DNTM) which is suited for parallel processing is designed by utilizing of the potential characteristic of parallel of SI algorithm. Three parallel computing models are designed based on the DNTM structure on the Drawning SMP Parallel Clustering Server and PC Cluster. The Parallel GA, Parallel PSO algorithm and Parallel QPSO algorithm are implemented by using MPI, OpenMP and MPI+OpenMP library. These three parallel algorithms are used to solve two kinds of nonlinear optimization problems. The results show that parallel QPSO algorithm has better performancesSecondly, the image registration based on maximization of mutual information algorithm and Powell algorithm is analysised. QPSO algorithm and parallelized QPSO algorithm are used as an optimization strategy in the image registration.Finally, two key steps in the process of image registration are improved by introducing a new similarity metric and a new hybrid searching strategy (by hybridizing NTPara-QPSO algorithm and Powell method). Based on the two improvements, the registration on the MR image and CT image are completed. Experimental results show that the hybrid algorithm has reached sub-pixels level and has better real time property.更多还原