Font Size: a A A

Research On Medical Image Registration Based On Information Theoretic Measure

Posted on:2017-02-03Degree:DoctorType:Dissertation
Country:ChinaCandidate:B C LiFull Text:PDF
GTID:1108330488957733Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Medical image registration, which plays very significant role in medical image fusion, rumor growth monitoring, image-guide surgical treatment, radiotherapy planning, and so on, is one of the most important technology in medical image processing. A variety of complementary information is fused by medical image registration, to provide more reliable information for doctors. Therefore, it is crucial to investigate medical image registration.Medical image registration has been attracted the attention of numerous researchers in the past many years, among which information theory-based image registration is the widely-used method and has been attracted the extensive attention, because it does not depend on the differences of gray values, without need of the preprocessing operation, such as feature extraction and segmentation, etc. In this thesis, several related issues of information theory-based registration technology are analyzed systematically, and some improvements and prefections are conducted based on the previous research works. The research works and contributions of this paper are mainly outlined as follows.(1) Medical image registration based on Jensen Arimoto divergenceBecause the Shannon entropy does not take the correlation of two independent random variables, while the non-extensivity of Arimoto entropy allows it considering this correlation. Thus, a similarity measure based on Arimoto entropy is proposed and used for medical image registration. In this method, rigid transformation and free-form deformations are employed as the spatial transformation model. In addition, in order to smooth the deformation in nonrigid registration, the bending enegy penalty is introduced as the smoothness term. Combining the dissimilarity measure and smoothness term, the objective function is constructed. Simultaneously, the Parzen window based on B-spline is adopted to estimate the joint probability distribution and the analytical derivative of the objective function is derived. Then, the limited-memory BFGS optimization scheme is exploited to obtain the optimal solution. The rigid and nonrigid registration experiments of simulated and clinical data demonstrate the accuracy results by using the method based on the Jensen-Arimoto divergence.(2) Histogram estimations method based on continuous image representation and its applications in medical image registrationTo relieve the influence of the number of bins and kernel function on the traditional histogram estimation, a strategy based on continuous image representation is presented to estimate the joint histogram of the images to be registered. Sequentially, the joint probability distribution is calculated and the required joint probability distribution is applied to compute the JAD similarity measure, and the measure is used to carry out medical image registration as the registration criterion. The joint histograms of 2D and 3D images are deduced by employing the continuous image representation. Additionally, in order to improve the execution efficience of histogram estimation method based on continuous image representation and reduce the grid effect in image registration, the fast continuous histogram estimation method is introduced by combining the random sampling theory. The rigid experimental results of 2D and 3D images illustrate that this approach accelerates the registration process compared to the continuous histogram estimation algorithm without incorporating random sampling while preserving registration accuracy; and the fast continuous histogram estimation method can achieve the higher registration accuracy than the simple histogram estimation method and Parzen window algorithm.(3) Medical image registration based on gradient distributions distanceThe classical approaches based on information theory does not consider the spatial information. To deal with this problem, a medical image registration method based on gradient distribution distance is proposed. In this method, the probability distributions of reference image and float image are estimated firstly, and then the KL (Kullback-Leibler) divergence is exploited to calculate the distance between two gradient probability distributions. Next, the gradient distribution distance is used as a penalty term of objective function so that the gradient distribution of float image is close to that of reference image. What is more, a normalized measure is defined as the dissimilarity term in terms of the property of JAD, and the ultimate objective function is constructed through incorporating the dissimilarity measure, the smoothness term and gradient distribution term. To obtain the analytical derivatives of the object function, in this method, the Parzen window approach is also applied to estimate the probability distributions, and then the gradient distribution distance is computed. The nonrigid registration experiments of clinical data indicate that the method based on gradient distribution distance can acquire the accuracy registration results.(4) Multimodal medical image registration based on structural image representationDemons algorithm reveals the inferior registration results in addressing multimodal medical image registration, for this reason, a diffeomorphism Demons registration method based on structural image representation is presented. The entropic images of reference and float images are calculated by Arimoto entropy, and they represent the spatial information of two images to be registered. Then, the diffeomorphism Demons algorithm is employed to register the two entropic images. Additionally, the update of displacement field between two registered images is obtained by optical flow equation, and the final displacement field is acquired with iteration approach. In this proposed method, two different modalities of the images to be registered are converted to the third kind of modality by employing the structural information, consequently, the registration of multimodal medical images is turned into the mono-modality problem. The experimental results show that the strategy based on Arimoto entropic image can achieve the higher registration accuracy compared to the diffeomorphism Demons algorithm and Demons method based on Shannon entropy.
Keywords/Search Tags:Mediccal image registration, Jensen Arimoto divergence, freee-form deformations, continuous image representation, gradient distribution distance, structural image representation, diffeomorphism Demons algorithm
PDF Full Text Request
Related items