Research On Intensity-Based Medical Image Rigid Registration Methods

Medical image registration is an important technique in the field of medical image processing, and is becoming more and more important for clinical diagnosis and treatment. Modern medical research usually requires integrated analysis of multiple images to get more information. A fundamental problem in medical image integrated analysis is that the images should be perfectly aligned, one essential aspect thereof is image registration, i.e., recovering the geometric relationship between corresponding points in multiple images of the same scene. Considering the registration robustness, precision and automation, the intensity-based registration methods are popular. This dissertation is devoted to the study of intensity-based rigid registration methods. The main work is summarized as follows:1. The sum of squared intensity differences between iinages is a simple and widely used registration function (registration measure) for monomodality image registration. In this dissertation, we respectively introduce a new totally structured secant method (NTSSM) and a new evolutionary strategy to optimize the registration function, and compare them with several methods (damping Gauss-Newton method, Powell method, simulated annealing method, classical evolutionary strategy—(μ＋λ)-ES) that are frequently used to optimize this function. It should be noted that unlike the Gauss-Newton method and its other improved methods discarding the second-order information of the Hessian matrix of objective function, the NTSSM method approximates the second-order information in an accurate way; and the update matrix remains positively definite during the computation. Registration results show that the NTSSM method can register images with higher precision and speed.2. Till now, mutual information-based registration method (MI) has been accepted by most researchers. MI measures the degree of dependence of two images by measuring the distance between their joint distribution and the product of their marginal distributions associated to the case of complete independence by means of Kullback-Leibler divergence (KLD) measure. In this paper, we introduce the use of the FDOD function to measure the aforementioned distance. Compared to the KLD measure, FDOD function has more appealing mathematical properties. Furthermore, inspired by the normalized mutual information (NMI), we propose a normalized FDOD function (NFDOD) to measure the distance. Experiments show that both FDOD function and NFDOD function are feasible in image registration, and the four methods (MI, NMI, FDOD, and NFOD) have similar registration precision. More importantly, NFDOD can register images successfully with less number of function evaluations than other three methods.3. Mutual information-based registration method usually obtains the joint probability distribution of two images by simply normalizing their joint histogram. Interpolation techniques are required during this process. But existing interpolation techniques will cause 'interpolation artefacts' which can result in local minima (or maxima). In this paper, we propose a new histogram estimation scheme which uses an exponential function to update the histogram continuously, and we derive the analytic expressions of gradient of the mutual information function derived from the new estimation scheme. Experiments show that the new histogram estimation scheme can significantly reduce the 'interpolation artefacts' and improve the registration robustness and precision. In addition, we suggest a hybrid normalized mutual information which uses the new proposed histogram estimation scheme for its smoothness and the classical histogram estimation scheme for its high speed in a complementary manner. The hybrid normalized mutual information method improves the registration robustness and speed efficiently.4. As the use of the registration package spreads, the number of the aligned image pairs in image databases (either by manual or automatic methods) increases dramatically. Any two images of the same or different acquisitions are aligned when the distance from their observed joint distribution to the corresponding expected distribution is minimized. In this paper, the distance is measured by the Tsallis divergence measure (TDM) for it is non-logarithmic. Experimental results show that, compared with the classical Shannon mutual information and Tsallis mutual information, our proposed method is computationally more efficient without sacrificing registration accuracy.