Sparse Representation And Digraph Based On Groupwise Image Registration Of Magnetic Resonance Brain Images

The goal of groupwise image registration(GIR)is to align all images from an image group to a common space.By doing this,same anatomical structures of different subjects have the same location.Currently,GIR methods are one of most popular techniques in medical field,for example,to provide doctors with efficient information for clinical diagnosis,to monitor the recovery of postoperative organs and to detect the functions of brain regions.Since high computational efficiency and registration accuracy,the graph based GIR(G-GIR)has been intensively studied.Clearly,image similarity measurement plays an important role in G-GIR.and image intensity based pairwise image similarity measurement(IPIS)is widely used in traditional G-GIR methods.However,IPIS only uses local information from the image group(i.e.,information of two images only),making the image manifold cannot be accurately estimated by the resulting graph.In traditional G-GIR methods,the deformation field registering each non-root image to the root image is obtained by composing deformation fields resulted from registration of adjacent images along the shortest path of each non-root image.Most G-GIR methods use static graph,where the graph built based on original images is never changed during the GIR process.The final registration accuracy of this method is not satisfactory for images that have obviously differences between images.In order to solve the above problems,the G-GIR method is improved in order to achieve more accurate registration results.Particularly,the tasks of this work are listed below:(1)This thesis introduces the related technologies of pairwise image registration,such as spatial transformation model and interpolation technique,briefly describes the evaluation criteria of image registration algorithm,and systematically describes the basic process of G-GIR methods,magnetic resonance brain images preprocessing methods and similarity measure between images.Compared with direct GIR method,the G-GIR method is better.(2)In this work,to solve the problem of inaccurate distribution manifold of images,which evaluated by the similarity between images using image grayscale information,an image manifold based global image similarity measurement using sparse representation is applied in G-GIR instead of Image intensity based pairwise image similarity measurement.Based on the calculated image similarities,a directed graph(digraph)can be built.Since global information is used,the estimated image manifold could be more accurate.This work register each non-root image and registered images to the father images along its shortest path subsequently until reach to the root image.By doing this,registration error accumulation can be solved.The results showed that:comparing with local information based G-GIR,the new algorithm is significantly improved in the registration accuracy.(3)In view of the large difference between the images in the image set,the traditional G-GIR methods based on static graph has its limitations.In this work,the graph is dynamically evolved.Specifically,an iterative process is applied,and in each iteration,non-root images are registered to their corresponding parent image according to their shortest path in the graph.Then based on the registered images,a new graph is built and sent to the next iteration.The iteration is convergent until all non-root images are registered to the root image.As the iteration goes,images get similar to each other,making the resulting graph closer to the true image manifold,and the registration can be enhanced.By comparing the G-GIR methods based on the static graph,the algorithm registration accuracy better.