Research On Key Technology Of Medical Image Registration

Medical image registration, a crossing research topic of information processing, computer image technology and modern medicine, has already had a wide range of application in clinical diagnosis, treatment, and preoperative plan. However, actual application indicates that it still faces many difficulties and challenges at various aspects such as the registration precision, speed, automation and robustness. Therefore, this dissertation summarizes the related concepts, methods and theory of medical image registration, focusing on key technologies in image registration process. To the drawback of some existing methods, improved or new differente methods are proposed. Some achievements as following are obtained.1. Aiming at these shortcomings of automatic landmarks extracting method such as methods based on area similarity and diatance map and etc. in medical image registration, an automatic landmarks extracting method based on shape matching is proposed. First, the two MR images to be registered are respectively segmented to extract their skull boundaries and cerebrospinal fluid boundaries and then shape context is applied to get landmarks on those boundaries and establish the corresponding relationship of landmarks between two images. Given the shape complexity of white matter, the shape context method is then improved with uniformly sampled points replaced by fewer modulus extreme points of wavelet transform on the white matter boundaries. This kind of landmarks can better reflect the shapes using fewer points and then the time for establishing their corresponding relationships is reduced. Finally, the landmarks selected with these methods are used to make B-spline surface interpolation, that is, image registration. Experimental results demonstrate that the landmarks gotten with this method can better reflect the geometry characters, have appropriate number and reasonable distribution in images; the B-spline-interpolation-based registration using the landmarks obtained with these methods has ideal accuracy and faster speed.2. Aiming at the drawback of registration method based on optical flow model that the assumption of constant brightness at a point makes it only used for the registration between signal-modal images, a model transformation method which use exact histogram specification is proposed, meanwhile, registration method based on optical flow model is also improved so that if the difference between two images to be made registration is large the registration result could also be ideal. In this method, landmarks which reflects the structural feature of images is taken using the above shape matching approach, and then additional external-force is constructed for optical flow model to get more ideal registration parameters. Experimental results demonstrate that the improved method can realize the accurate registration of multi-modality time-sequence MR images between which there is large difference.3. A novel method combining feature constraint with multilevel strategy to improve simultaneously the registration accuracy and speed is proposed for non-parametric image registrations. To images between which the local difference is large, integrating feature constraint constructed with local structure information of images into objective function of image registration improves the registration accuracy. When applying feature constraint under multilevel strategy, parameter searching is prevented from entrapped into local extremum by using the optimization result on coarser levels as the starting points on finer levels; meanwhile traditional optimization methods but not intelligent optimization algorithms which consume more time can find the accurate registration parameters on finer levels, so registration speed is improved. Experimental results indicate that this method can finish fast and accurate registration for images between which there exists large local difference.4. Because of the local maxima of mutual information caused by image noise and interpolating, B-spline-based non-rigid registration of medical multimodal images which use mutual information as similarity measure may not find the best registration parameters when traditional optimization is taken as the search strategy, so particle swarm optimization (PSO) is studied. In order to prevent to be trapped in local optimum, the registration parameters obtained by LBFGS(Limited-memory Broyden-Fletcher-Goldfarb-Shanno) optimization are utilized to make initial particle swarm at first, then the multi-objective optimization and cross-mutation strategy are combined in PSO optimization. Registration experiments confirm the validity of the proposed method.In addition, during working on PhD, I also participated in subject about the non-stationary signal compression. After some concrete work, I put forward a novel method to compress non-stationary signals using compressive sensing. This method constructs the sparse basis for non-stationary signal and the number of samplings is decreased with compressive sensing. The signal reconstructed with orthogonal matching tracking method is very near to the original signal in time domain and frequency domain as well as the time-frequency domain.Computer simulation shows its validity.